Load-adaptive networking for energy-efficient wireless access

ARTICLE IN PRESS

JID: COMCOM

[m5G;June 12, 2015;7:19]

Computer Communications 000 (2015) 1–9

Contents lists available at ScienceDirect

Computer Communications journal homepage: www.elsevier.com/locate/comcom

Load-adaptive networking for energy-efficient wireless access✩ Nico Bayer a,∗, Karina Gomez b, Cigdem Sengul c, Dirk von Hugo a, Sebastian Göndör d, Abdulbaki Uzun d a

Telekom Innovation Laboratories, Berlin, Germany CREATE-NET, Trento, Italy c Oxford Brookes University, Oxford, United Kingdom d Telekom Innovation Laboratories, TU Berlin, Service-centric Networking, Berlin, Germany b

a r t i c l e

i n f o

Article history: Received 21 March 2014 Revised 12 May 2015 Accepted 13 May 2015 Available online xxx Keywords: Wireless networks Green networking Load-adaptive network

a b s t r a c t Energy-efficient operation is essential for mobile network operators to meet the growing demand for higher data rates while managing rising operating costs. Here, the main challenge is to guarantee the quality of user experience whilst saving energy. This challenge demands adaptive algorithms that enable a load-aware network operation that dynamically configures different network elements according to user needs. To this end, in this paper, we present an adaptive and context-aware power management framework for networks composed of different radio access technologies. We implement and evaluate our framework in an indoor and outdoor testbed. The experimental results confirm that significant energy can be saved in practice by efficiently adapting resources to the actual traffic demand.

1. Introduction One of the most urgent challenges in the new century is reducing the high energy consumption. As mobile Internet becomes more prevalent, energy consumption related to communication networks also increases. In 2010, the global electricity and diesel energy consumption by all mobile networks was approximately 120 TWh, resulting in $13 billion energy costs and, is responsible for 70 Mt CO2 emissions [1]. This is not surprising, as, typically, mobile networks are over-provisioned to provide the best connectivity to the users. However, the actual demand varies considerably based on context (e.g., time and location). To save energy in mobile networks, this mismatch in demand and provisioned resources needs to be addressed. An attractive solution proposed in literature is to adaptively re-configure the mobile network elements (e.g., switching base stations to lower energy consumption modes or off) to meet current capacity demands. In this paper, we evaluate the potential of such an approach based on a context-aware energy management framework implemented in an indoor and outdoor testbed. ✩ This research was partly funded by the Communicate Green project (Grant number: 01ME11010) within the IT2Green Initiative of the BMWi in Germany (www.communicate-green.de). ∗ Corresponding author. Tel.: +491605366062. E-mail addresses: [email protected] (N. Bayer), [email protected] (K. Gomez), [email protected] (C. Sengul), [email protected] (D. von Hugo), [email protected] (S. Göndör), [email protected] (A. Uzun).

© 2015 Elsevier B.V. All rights reserved.

One of the major challenges in context-aware energy management is to provide timely and accurate network re-configurations as the user demand varies. In this work, we consider a load-aware energy-saving framework, which takes into account that radio access networks are heterogeneous in structure (e.g., cell size) and technology (e.g., using different wireless technology standards [2]). This work builds on our earlier works [2–6] which introduce the conceptual framework of a load-adaptive network. In this paper, we provide further analysis on load-adaptive network design (Section 3), which includes a discussion about the main parameters and costs that influence energy-efficiency of a load-adaptive system. This is essential to ensure that the cost of energy management mechanisms do not negate energy savings. The proposed framework is made up of several components including an energy optimiser, and a context manager, context collection agents and a controller (see Section 4 for details). We fully integrated our energy optimiser, Morfeo [4] and context manager [3] into this framework along with other components. We ran extensive tests using a testbed with a total of 19 outdoor and indoor wireless access devices (WADs) and 20 users. We also used two types of WADs with different energy consumption, capacity and coverage characteristics to understand their impact on energy savings and load balancing. Our results demonstrate that significant energy savings are possible, but depend on hardware and software characteristics of the WAD as well as the required network coverage and actual traffic load. In summary, we make the following contributions in this work:

http://dx.doi.org/10.1016/j.comcom.2015.05.004 0140-3664/© 2015 Elsevier B.V. All rights reserved.

Please cite this article as: N. Bayer et al., Load-adaptive networking for energy-efficient wireless access, Computer Communications (2015), http://dx.doi.org/10.1016/j.comcom.2015.05.004

JID: COMCOM 2

ARTICLE IN PRESS

[m5G;June 12, 2015;7:19]

N. Bayer et al. / Computer Communications 000 (2015) 1–9

• Design and implementation of a context-aware energy management framework which provides load-aware network operation. • Design of experiments to measure the system parameters that influence the implementation of the energy saving framework in a real testbed. • Extensive experiments in indoor and outdoor testbeds to evaluate different traffic conditions (e.g., no traffic, constant traffic, increasing traffic). • Analysis and discussion of the limitations of energy-savings using a power-off strategy. The remainder of the paper is structured as follows. We present the state of the art in Section 2, and in Section 3, discuss the design of a load-adaptive network. Sections 4 and 5 present our implementation and experimental evaluation, respectively. Section 6 concludes the paper. 2. Greening ICT networks: state of the art Greening ICT networks has become a central theme in research and standardisation activities due to expected high increase in energy consumption [7–9]. There is extensive literature on energy saving solutions in mobile networks including 2G, 3G and LTE advanced, and also Wireless LANs (WLANs) and other short-range radio technologies. The standard use-cases consider saving energy (1) at a single Access Point (AP) of a given technology, (2) among APs of a given technology, and (3) across different networks of different technologies [10]. Nevertheless, the majority of the proposed solutions focus on single technology rather than a heterogeneous network. Also, common solutions typically either trigger low-power modes during periods with no network traffic, or use power control for cell zooming and exploit coverage and energy trade-offs. In the following, we present further details on these two approaches. Low-power modes: Using low-power modes to save energy is well-investigated [11–14]. Here, with low-power mode, we mean both a sleep mode, which deactivates different parts of a WAD, as well as an off mode, which shuts down the entire device. The main goal of using a low-power mode is to reduce the number of active base stations or sectors under low traffic conditions. If the overhead of off/sleep and on transitions are acceptable, significant energy savings become possible [15]. The OPERA-Net (Optimising Power Efficiency in mobile Radio Networks) project [16] shows that 30% energy saving is achievable without affecting user satisfaction when a sleep mode is used to turn off sectors in a 3G network. Similarly, the E3 (Endto-End Efficiency) project [17] has quantified the gains from switching off sectors as 20–40%, while maintaining an acceptable grade of service for different traffic types. The TREND (Towards Real Energyefficient Network Design) project [18] proposes using sleep modes in home femtocell networks, where long under-utilised periods are the norm, and hence, high energy savings can be obtained without affecting service quality. The EARTH (Energy-Aware Radio and network Technologies) project [19] proposes several deployment and network management strategies, and sophisticated hardware and software techniques to achieve overall energy savings of 70%. Among these, the cell DTX (discontinuous transmission), which allows putting a transceiver to a low-power mode, is shown to be the most efficient approach when combined with bandwidth adaptation (e.g., adapting the resource blocks in an LTE subframe according to the traffic demand). In addition, it may also be possible to turn off entire networks [20,21]. In [20], using heterogeneous 2G/3G networks of the same operator, an optimal strategy powers on/off an entire system (2G or 3G) for high or low traffic scenarios, respectively. In [21], significant savings are achieved if operators enable cooperative roaming and switch off networks under low load. The main challenge, as these studies confirm, is to offload the current traffic to active cells

and avoid coverage holes. In [22], re–arranging the user–cell association is shown to achieve 50% energy savings. However, this result serves as an upper bound, as complete knowledge of user locations is assumed. In [23], the temporal-spatial user traffic diversity is exploited to turn off 3G base stations. A network trace-based evaluation shows gains up to 52.7% savings in a dense area, and up to 23.4% in a sparse area. Savings are also more significant during night time, e.g., up to 70%. However, even during daytime, 20–40% savings are possible by exploiting temporal-spatial traffic diversity. Similar strategies also exist for WLANs [24–27]. Cell zooming: Cell zooming via transmit power control requires determining the optimal transmission power for each base station while maintaining good QoS. Here, we consider two approaches: (i) adjusting the transmission power to obtain the Signal to Interference plus Noise Ratio (SINR) that achieves the desired network capacity with the current modulation coding scheme, and (ii) adjusting the transmission power accepting a degradation in SINR, which may be salvaged by more robust modulation coding schemes. In [28], the optimum cell size based on different base station technologies, data rates, and traffic demands is investigated and a two-level scheme is shown to achieve up to 40% energy savings. The EARTH project also considers power control as an attractive solution for high load scenarios and in combination with cell DTX. Cell zooming is proposed to be used in combination with sleep modes to alleviate coverage holes resulting from switching off base stations, and to extend coverage. However, in [29], it is shown that this may be insufficient compared to deploying smaller but more cells. Similarly, [30,31] propose a heuristic approach to adopt the on/off activity of APs and corresponding transmit power in accordance with number and location of active users. The main goal is to strike a balance between computational complexity and accuracy to enable energy savings. Power-bandwidth optimisation techniques are investigated in [32], where authors also discuss practical trade-offs associated with the implementation of the energy efficient schemes. The paper concludes that energy efficient techniques provide considerable power savings even accounting for realistic system parameters and channel environments. 3. Designing a load-adaptive network The main goal of a load-adaptive network is to configure itself to satisfy the actual capacity demand and also to reduce the energy consumption of the network. To this end, we first identify the main system and design parameters, and next, discuss network configuration decisions that affect the capacity and energy consumption of a network. Finally, we propose a functional architecture that takes into account these parameters and decision making procedures to achieve a load-adaptive operation. 3.1. Design parameters and decision mechanisms The context and deployment-based parameters that we take into account are summarised in Table 1. In our design, we include the delay and energy cost of network re-configuration (i.e., τ and Erec ), and context collection delay (i.e.,  ). These costs are zero only in the ideal case and, it is not possible to re-configure the network continuously and in real-time, but at specific points in time and with a configuration that needs to be valid for a minimum time period (i.e., δ in Table 1). This also requires mechanisms that can predict the capacity demand during δ . In addition, some over-provisioning is required to be able to handle traffic bursts. In the case that the actual demand significantly differs from the predicted demand, additional mechanisms are required to detect and correct those situations. It is clear that a re-configuration only makes sense if the saved energy is greater than zero (Es > 0).

Please cite this article as: N. Bayer et al., Load-adaptive networking for energy-efficient wireless access, Computer Communications (2015), http://dx.doi.org/10.1016/j.comcom.2015.05.004

JID: COMCOM

ARTICLE IN PRESS

[m5G;June 12, 2015;7:19]

N. Bayer et al. / Computer Communications 000 (2015) 1–9

3

Fig. 1. Example network deployment and neighbour relation table.

Table 1 Parameters of a load-adaptive system. Context-based parameters Current context (C) Current load (V) Context collection delay ( ) Context collection error ( ) Deployment-based parameters Capacity demand (CD ) Re-configuration delay (τ ) Re-configuration energy (Erec ) Re-configuration granularity (G) Context collection E (ECoCo ) Tunable operation parameters Re-configuration interval (δ ) Provided capacity (CP ) Over-provisioning factor (OP) Overload thresh. (Tover ) Underload thresh. (Tunder )

Current load, number of users and their location etc. Current IP-layer data to be transported. Signalling and processing delay for context collection. Error in context collection and prediction mechanisms. V translated to capacity demand at the physical layer. Software and hardware delays of re-configurations. Energy used in activating a configuration. The granularity in which a configuration affects capacity. Energy required for context collection.

has three sectors with one carrier per sector. Fig. 1 shows the neighbour relation table. Neighbour relations may be asymmetric due to different technologies having different coverage as they use different frequency bands, transmit powers, antenna tilts etc. As expected, mobiles connected to the underlay cell see the overlay cell with a very high probability, and conversely, the mobiles connected to the overlay cell see the underlay cell with a lower probability. In this context an overlay cell refers to a cell that provides coverage (e.g. Global System for Mobile Communications (GSM)) while an underlay cell refers to a cell that provides capacity (e.g. Long Term Evolution (LTE)). Based on the neighbour relation table, the neighbour relation graph is constructed to decide which nodes to include in a re-configuration. For instance, before switching off a cell of an underlay network, the neighbour cells are checked for overloading. If there is overloading, the load balancing algorithm should handover mobile devices from the overloaded cell to this underloaded cell. If this is not the case, the cell can be switched off. 3.2. System architecture

Time between re-configurations (δ > τ ). Capacity provided by the current configuration (CP > CD ). Excess capacity provided by the current configuration. The maximum load threshold used in load-balancing. The minimum load threshold used in load-balancing.

Given these requirements, a load-adaptive system may require a central entity, which carefully decides re-configuration times based on the current context. Decision mechanisms in a load-adaptive network need to answer two questions: (i) when to re-configure the network and ii) which WAD to involve in re-configuration. As a result of these decisions, load balancing mechanisms may need to run to ensure that all user demand is met. In order to decide the reconfiguration times, we use a threshold-based policy, where we denote Tunder and Tover as the two thresholds to represent the cases when a network is underloaded and overloaded, respectively. The decisions on which wireless devices to include in the configurations can be taken at base station, cell site, or network level. For all three cases, we use a neighbour relation table to assist configuration decisions. This table is used to record the dependencies between different cells in a network based on live network statistics gained from measurements. These measurements report information about neighbouring cells seen by a mobile device while connected to its serving cell. Based on these statistics, coverage overlaps of cells are determined. While such measurements can be collected by drive tests performed by the mobile operator, several standardisation activities are also discussing the possibility of measurements performed directly on mobile devices [33]. To illustrate the use of neighbour relation tables, Fig. 1 depicts an example case with two BSs of different technologies (one in grey and one in white) deployed at the same site. Each BS

Our system architecture consists of two building blocks: (i) the Context Management Framework (CMF) as proposed in [3], and (ii) the Energy Optimiser (EO) (see Fig. 2). The CMF is responsible for collecting, storing and processing contextual data in the network. This data is used to compute a model of the current state of the network and its elements. Based on this information, the EO re-configures the network to match the capacity demand of the users with minimal energy consumption. 3.2.1. Context management framework The context management architecture integrated into our framework was presented in [3]. In this section, we give a brief summary of its main elements: Context Source, Context Collection Agent (CCA), and Context Manager. The CCA runs on a Context Source for acquiring contextual data provided by this source. This data can be collected from various sensors (e.g., GPS or accelerometer), device drivers (e.g., Wi-Fi network adapter), or any third-party service (e.g., Web APIs providing information about the weather or upcoming events). In this paper, we refer to Context Sources as WADs (e.g., AP or BS) providing data such as the traffic load, the number of connected users, or the power consumption. The Context Manager (CM) is the central entity and is responsible for aggregating and processing the data collected by multiple CCAs. Each CCA registers itself by publishing what type of context data it can deliver. This way, the CM can assemble a list of all types of context information provided by the system. All context data received by the CM is stored in an internal Context Database, which is accessible via the CM-EO-A interface1 by other entities such as the EO. 1 Notice that this notation (CM-EO-A) denotes an interface which provides two distinct types of information, signalling and traffic, between the Context Manager and the Energy Optimiser.

Please cite this article as: N. Bayer et al., Load-adaptive networking for energy-efficient wireless access, Computer Communications (2015), http://dx.doi.org/10.1016/j.comcom.2015.05.004

JID: COMCOM 4

ARTICLE IN PRESS

[m5G;June 12, 2015;7:19]

N. Bayer et al. / Computer Communications 000 (2015) 1–9

Fig. 2. Functional architecture.

3.2.2. Energy optimiser EO runs an energy optimisation algorithm, which takes information from the Context Database (via the CM-EO-A interface) and calculates an energy-efficient network configuration to serve the user demand. Hence, EO makes the necessary decisions to choose operation modes, and schedules the actions to be executed by the Controller. EO may run different energy optimisation algorithms. We give an example algorithm in Section 4.2, called morfeo. The Controller, based on EO commands, triggers the necessary actions, e.g., changes transmit powers, switches on/off cells, etc. For this purpose, the Controller is connected to Configuration Agents (CAs), which are modules able to execute the new configurations. The CAs can be, for instance, an operation support system (OSS) or a remote power switch (e.g., Energino [34]).

(see Fig. 2). In this paper, WADs support the following four modes of operation:

4. Implementation

Based on these modes, we summarise the operation of morfeo with three steps: initialisation, reactive updates, and correction. In the initialisation step, using the neighbour relation graph, δ , CP , CD and OP (described in Table 1) morfeo groups WADs into three different groups. Power save candidates can reduce the energy consumption if switched to Off-Mode or Low-Power Mode. Head candidates can take over the load of other WADs (e.g., due to overlapping coverage). As these nodes provide coverage, they are not switched off. Finally, special candidates need to be always in AM and are excluded by the network operator from the energy optimisation process. This might be relevant in special situations, e.g. in which availability is more important than energy efficiency. In the reactive updates step, morfeo evaluates several conditions for each WAD based on network current load V. These conditions are

4.1. Context collection In our testbeds, we implemented CCAs for Wi-Fi APs with multiple interfaces to collect context data related to network and energy use. More specifically, the Wi-Fi CCA tracks the number of clients, total traffic, per-client traffic, and the energy consumption of the device. The implementation consists of a main module and additional modules for collecting as well as sending different context data. This way, data sources can be added and removed easily from the CCA. The main module is built as a time-based job scheduler that registers the Wi-Fi WAD with a unique identifier (cca_key) at the CM and keeps control of the modules. If the application is stopped for any reason, e.g., due to operating system upgrades, a de-registration is performed automatically by the CM. Collected information is stored locally on the machine using a MySQL database server and can be requested independently by the module responsible for sending the data to the CM. For error tracing and debugging, a history of all CCA events is created by logging every operation of a module regardless of whether it succeeded or failed. This ensures that missing data due to a send failure will be resent automatically. In addition, for each module, an individual execution interval can be set, allowing for a fine-grained configuration. 4.2. MORFEO: energy-saving decision algorithm We present morfeo2 as an example energy-saving algorithm located in EO [4,6]. Morfeo interacts with the CAs, through a controller 2 Morfeo is proposed originally in [4], and presented in this article for the sake of completeness.

• Full–Power Mode (PM): The WAD is on and operating with the maximum transmission power (txPower) level. This mode provides full coverage and capacity. • Active Mode (AM): The WAD is on and operating with the default txPower level, which might be lower than the maximum txPower. • Low-Power Mode (LM): The WAD is on but one or more of its sectors/interface are off. Note that this is different than sleep modes typically implemented in the client-side wireless interfaces. In our implementation, the switching on and off of this mode is controlled by energy optimiser, rather than pre-defined sleep schedules. • Off Mode (OM): The WAD is off.

• Zero condition: A WAD does not have any associated client (V = ∅). Power save candidates in zero condition are switched to OM. • Idle condition: A WAD has clients but none of them are sending or receiving (V = 0). Power save candidates in idle condition are switched to LM. • Active condition: A WAD has clients with traffic (V > 0). In this case, load balancing is triggered and users are handed over to the head candidates, if it is possible. WADs in this condition are added back to the set of power save candidates to check the possibility of switching to LM or OM. • Head condition: A WAD takes over the load of the others that are switched to OM or LM. Head candidates in head condition are switched to PM only when coverage adaptations are needed (CP > CD with OP > 0). Before deactivating cells, morfeo also evaluates if the active cells have enough capacity (CP > CD ) for supporting the incoming traffic

Please cite this article as: N. Bayer et al., Load-adaptive networking for energy-efficient wireless access, Computer Communications (2015), http://dx.doi.org/10.1016/j.comcom.2015.05.004

ARTICLE IN PRESS

JID: COMCOM

[m5G;June 12, 2015;7:19]

N. Bayer et al. / Computer Communications 000 (2015) 1–9

(a)Indoor WADs (2nd floor)

(b) Indoor WADs (3rd floor)

5

(c) Outdoor WADs (top)

Fig. 3. Network topology for testing scenarios.

RTT

100

250

80

Switch OFF

60

E

200

E

60

300

rec1

150 40

rec2

20 0 0

100

E

50 10 τ

20

(AM−OM)

30

40

50

60

70

80 90 τ(OM−AM)

0 100 110 120 130 140 150 160 170 Time [s]

(a) Measurements for an outdoor WAD

300 RTT

Current Consumption

Switch ON

200

rec1

150

40

E

100

rec2

20 0 0

250

Current [mA]

Current Consumption

Switch ON

Switch OFF

RTT [ms]

RTT [ms]

80

Current [mA]

100

50 10 20 30 τ(AM−OM)

40

50 τ

60

70

80

90

0 100 110 120 130 140 150 160 170

(OM−AM)

Time [s]

(b) Measurements for an indoor WAD

Fig. 4. Measurements of τ and Erec for different WADs of the testbed considering off mode and active mode.

from the deactivated cells. In this way overload of the active cells is avoided. Then morfeo applies the necessary changes and switches to the Correction step after δ > (τ +  ) (see Table 1). In the correction step, morfeo analyses CP and CD based on the statistics collected from CM. If CD is greater than CP or the quality of service (QoS) constraints are not satisfied, the WADs that are in lowpower modes are switched to AM. Otherwise, morfeo calls Reactive Updates to discover if more WADs could be put to low-power modes. If the analysis indicates further energy savings, the Reactive Updates step evaluates the changes in operation modes and goes back to Correction step. Finally, morfeo continues with reactive updates and correction steps in a cycle. 5. Evaluation We evaluate the proposed framework in a testbed, composed of four outdoor WADs, 15 indoor WADs and 20 clients deployed at Telekom Innovation Laboratories in Berlin, Germany. The testbed covers an area of approximately 9600 m2 , which is divided into four separate courtyards. Fig. 3 shows the network deployment: indoor WADs are deployed in the 2nd and 3rd floor while the outdoor WADs are deployed on the top of the building (shown as APs in white rectangles) serving 20 users distributed in the building (shown as STAs in yellow rectangles). The outdoor WADs are powered with 24 V DC power supply and are commercially available Saxnet meshnode III equipped with four wireless IEEE 802.11a/b/g interfaces working in AP mode. More details about the outdoor testbed can be found in [5]. The indoor WADs are powered with 12 V DC power supply and are commercially available Embedded Server with Intel Atom D525 CPU, 4GB RAM and two IEEE 802.11a/b/g interfaces. Overall this means that the scenario comprises 46 wireless interfaces. The outdoor and indoor WADs are connected via Ethernet to a wired backbone network. The rate adaptation algorithm has been set to the default auto, and the transmission power has also been left as the default value of 17 dBm (≈50.12 mW) for outdoor WADs and 15dBm (≈32 mW) for indoor WADs during all experiments.

5.1. System parameters calculation For our evaluation, the most important system parameters are

τ , Erec ,  and ECoCo (as discussed in Section 3.1). In this section, we

present how these values are determined for our testbed and hardware. As explained previously, switching between different modes has both an energy (Erec ) and a time (τ ) cost. To determine τ and Erec , we first consider the case of AM and OM switching. We define a minimum off–time (t∗ ) for which the energy gain from switching to OM compensates the energy waste in switching from AM to OM, and from OM to AM. Fig. 4 depicts the measurement results based on indoor and outdoor WADs. The outdoor WAD takes ≈4 s to turn off completely (τ(AM−OM ) = 4 s) and ≈110 s to be completely operational again (τ(OM−AM ) = 110 s) (see Fig. 4a). Since, the AM and switching mode powers are almost the same, t∗ ≈ 0. Therefore, once the WAD is in OM, morfeo does not evaluate the conditions that can lead to its wake–up before t ∗ + τAM−OM + τOM−AM , which is approximately 114 s. In Fig. 4b, the indoor WAD takes ≈4 s to turn off completely (τAM−OM = 4 s) and ≈51 s to be completely operational again (τOM−AM = 51 s). Here, t∗ is again ≈0, consequently morfeo does not evaluate the conditions for waking up the WADs before τAM−OM + τOM−AM , which is approximately 55 s. Note that the values of τ and Erec for outdoor and indoor WADs are different due to the use of different hardware and the POM value of the WAD is not equal to zero due to the additional power meter device needed for energy monitoring and implementation of the LM. The power meter devices are always on and able to communicate with the controller using an Ethernet interface [34]3 . Additionally, it is important to underline that the WADs cannot be powered off instantly as shown in the figure. This is due to the fact that WADs are powered off by software and not only by cutting the power to the WADs. Basically, when the EO decides to put a WAD in OM, it sends a command to the con3 The difference of the POM value between both indoor and outdoor WADs is due to the outdoor WAD having additional equipment for monitoring weather conditions, which is not switched off when the outdoor WAD is in Off Mode.

Please cite this article as: N. Bayer et al., Load-adaptive networking for energy-efficient wireless access, Computer Communications (2015), http://dx.doi.org/10.1016/j.comcom.2015.05.004

ARTICLE IN PRESS

JID: COMCOM

4.2 4.0 3.8 3.6 3.4 3.2 3.0 2.8 2.6 2.4 0

Power Consumption [W]

N. Bayer et al. / Computer Communications 000 (2015) 1–9

Power Consumption [W]

6

[m5G;June 12, 2015;7:19]

Power Consumption wlan0 up wlan1 up

wlan0 up wlan1 down wlan0/wlan1 down

τ

(LM−AM)

0.5

1

1.5

2

2.5

3

and E

rec3

3.5

4

4.5 5 Time [s]

wlan0 down wlan1 up

τ

(AM−LM)

5.5

wlan0 down wlan1 down

and E

rec4

6

6.5

7

7.5

8

8.5

4.2 4.0 3.8 3.6 3.4 3.2 3.0 2.8 2.6 2.4

9

(a) Power consumed versus time

Power Consumption Confidence Interval

wlan0/wlan1 wlan0 up wlan1 down down

wlan0 up wlan1 up

(b) Average power consumed

34.533 ± 0.2613

ε from CA Database to Optimizer for 19 WADs 1

2

3

4

5

6

7

8

ε |ε|

9 10 11 12 13 14 15

Samples

(a) Measurements of 

1.860 Ecoco(t=1s)

1.836 1.812 1.788 1.764 1.740

CCA enabled

CCA disabled

(b) Indoor WAD.

Power Consumption [W]

37 36 35 34 33 32 31 30

Power Consumption [W]

ε [s]

Fig. 5. Measurements of τ and Erec for outdoor WAD of the testbed considering low-power mode and active mode.

4.200 Ecoco(t=1s)

4.152 4.104 4.056 4.008 3.960

CCA enabled

CCA disabled

(c) Outdoor WAD.

Fig. 6. Measurements of  and energy cost for collecting context information of users.

troller for turning off the WAD. The controller then (i) powers off the WAD by software, and (ii) when the WAD is off, the controller cuts off the power to the WAD using the power meter. This procedure is done to avoid damages to the hardware and software of the WADs. In addition, we experimented with different configurations for LM with outdoor WAD using two wireless interfaces. The results in Fig. 5 show that the minimum time required for saving energy to switch a WAD to LM from AM or vice-versa is τLM−AM = τAM−LM = 0. Therefore, there is no condition on the minimum time to spend in LM. Also, collecting the context information has energy (ECoCo ) and time ( ) cost. In Fig. 6 the measured values of  and ECoCo (t = 1 s ) are shown. In Fig. 6a, the average time  for collecting users and traffic statistics of the whole network is ≈34.533 ± 0.2613 s (the statistics are collected from each wireless interface of the 19 WADs). The ECoCo for a single indoor and outdoor WAD are depicted in Fig. 6b and c. Here, to measure ECoCo value, we compare the current draw of the WAD when CCA is enabled and disabled, respectively. It is important to remark that the obtained cost from the analysis of the various system parameters (i.e, OM-AM switching costs) are not intended to be general since they depend on different types of hardware, software and technology. However, these costs are presented with the goal of explaining how to measure these parameters. 5.2. Evaluation and discussion We ran a measurement campaign to evaluate the energy savings obtained by our solution. The power consumption of each WAD is measured using the EPS4 power meter, which logs power with a granularity of 0.1 W and a sampling period of 1 s. Table 2 shows the network topology used in our tests. The WADs are distributed over three floors, and there are six power save candidates and 13 head candidates (see Table 2 for their IDs, which match the IDs shown in Fig. 3).

4

Available at: www.gude.info.

Table 2 Network topology. Floor

WADs

Power save candidates

Head candidates

Users

2nd floor 3rd floor Top

9 6 4

2 (75, 60) 3 (15, 65, 69) 1(55)

7 3 3

6 6 8

Next, we discuss four different experiment scenarios. We ran each experiment four times for a duration of 1000 s, and present the results with their 95% confidence interval. - Association scenario (Scenario 1): Users are associated to the network, but do not generate traffic. The objective of this experiment is to evaluate energy savings in an idle network, which is optimal for saving energy (i.e., V = 0). Fig. 7 shows the average power consumption of the WADs when EO is on and off. The energy savings achieved are approximately 18% for outdoor network and 32% for indoor network (see the Fig. 7a). Since there is no network traffic, the Power Save Candidates remain constantly in OM (the power save candidates are marked in Fig. 7b with arrows). - Constant traffic (Scenario 2): In this scenario clients, associated to different WADs, generate traffic after EO is turned on and the network is completely reconfigured. So, the power save candidates are already in OM when the traffic is generated. Notice that the generated traffic is lower than the capacity provided by the current configuration (CP > CD ). This means that power save candidates remain in OM and there are no handovers to other WADs (head candidates). The goal of this experiment is to test the network performance when there is network traffic but no load balancing is required. The traffic is generated using the Iperf and emulates video traffic by sending UDP packets with a bitrate of 1.2 Mb/s during 600s. This corresponds to a video coding layer (VCL) for high resolution (320240@20). Eight senders upload one video to a central server and eight receivers download one video from a central server. The central server is

Please cite this article as: N. Bayer et al., Load-adaptive networking for energy-efficient wireless access, Computer Communications (2015), http://dx.doi.org/10.1016/j.comcom.2015.05.004

ARTICLE IN PRESS

JID: COMCOM

[m5G;June 12, 2015;7:19]

Power Consumption [W]

EO Off 4 3.5 3 2.5 2 1.5 1 0.5 0

EO On

18% Power Savings 32% Power Savings

Outdoor WAD

Power Consumption [W]

N. Bayer et al. / Computer Communications 000 (2015) 1–9

4.5 4 3.5 3 2.5 2 1.5 1 0.5 0

Indoor WAD

WAD in Off Mode Power Save Candidates

7

EO Off EO On

52 54 55 56 18 11 75 27 68 60 21 50 79 20 69 89 15 40 65 WAD ID

(a) Average Power

(b) Power Consumption per WAD

Power Consumption [W]

EO Off 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0

EO On

20% Power Savings 31% Power Savings

Outdoor WAD

Power Consumption [W]

Fig. 7. Scenario 1: average power consumption of the WADs when EO is on and off. Network users are idle (V = 0).

4.5 4 3.5 3 2.5 2 1.5 1 0.5 0

Indoor WAD

WAD in Off Mode Power Save Candidates

EO Off EO On

52 54 55 56 18 11 75 27 68 60 21 50 79 20 69 89 15 40 65 WAD ID

(a) Average Power

(b) Power Consumption per WAD

Fig. 8. Scenario 2: average power consumption of the WADs performing when EO is on and off. The network has users generating traffic (V = 0).

Table 3 Average of bandwidth and packet loss achieved for users. Scenarios

Packet loss uplink [%]

Bandwidth uplink [Mb/s]

2 (EO Off) 2 (EO On) 3 (EO Off) 3 (EO On) 4 (EO Off) 4 (EO On)

4.50 ± 1.50 4.60 ± 1.38 4.60 ± 1.42 8.60 ± 3.08 4.40 ± 2.45 6.50 ± 2.08

1.14 ± 0.04 1.13 ± 0.03 1.20 ± 0.03 1.00 ± 0.30 2.59 ± 0.04 2.20 ± 0.40

Packet loss downlink [%] 5.50 ± 1.50 6.00 ± 1.60 6.00 ± 1.60 10.00 ± 4.60 5.00 ± 1.00 8.00 ± 2.70

Bandwidth downlink [Mb/s] 1.13 ± 0.02 1.16 ± 0.01 1.19 ± 0.03 0.99 ± 0.10 2.78 ± 0.05 2.55 ± 0.02

Power Consumption [W]

EO Off 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0

EO On

19% Power Savings 30% Power Savings

Outdoor WAD

Power Consumption [W]

accessible via Ethernet from each WAD. Fig. 8 shows the average power consumption of the WADs when EO is off and on. The energy savings are similar to the previous scenario. Table 3 shows that the average bandwidth and packet loss achieved in uplink and downlink are similar for EO off and on. Hence, the same network performance is achieved while saving energy.

Indoor WAD

(a) Average Power

4.5 4 3.5 3 2.5 2 1.5 1 0.5 0

- Load balancing constant traffic (Scenario 3): In this scenario, each user starts generating traffic at the same time when the EO is carrying out network reconfigurations. So, the clients are initially associated with random WADs, but need to be handed over to head candidates if their WAD is a power save candidate. The goal of this experiment is to test network performance with load balancing. The traffic is generated as explained in Scenario 2. Fig. 9 shows the average power consumption of the WADs when EO is off and on. Table 3 shows that the network performance is affected, and there is a higher packet loss due to network reconfiguration. - Load balancing increasing traffic (Scenario 4): In this scenario, clients increase their traffic randomly (up to 3 Mb/s). The traffic is generated in the same way as in the previous scenarios, but we test network performance under more dynamic conditions. Fig. 10 shows the average power consumption of the WADs when EO is off and on. The energy savings are up to 9% for outdoor network and 25% for indoor network (see the Fig. 10a). In Fig. 10b the network power consumption versus the time is depicted, and shows that multiple network reconfigurations are necessary to serve the users. Since the

WAD in Off Mode Power Save Candidates

EO Off EO On

52 54 55 56 18 11 75 27 68 60 21 50 79 20 69 89 15 40 65 WAD ID

(b) Power Consumption per WAD

Fig. 9. Scenario 3: average power consumption of the WADs performing without and with EO enabled. The network users generating traffic (V = 0) and EO performs users load balancing.

Please cite this article as: N. Bayer et al., Load-adaptive networking for energy-efficient wireless access, Computer Communications (2015), http://dx.doi.org/10.1016/j.comcom.2015.05.004

ARTICLE IN PRESS

JID: COMCOM 8

[m5G;June 12, 2015;7:19]

N. Bayer et al. / Computer Communications 000 (2015) 1–9

EO On

9% Power Savings 25% Power Savings

55 Network Power [W]

Power Consumption [W]

EO Off 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0

50 45 40 EO enabled 35 30 0

Outdoor WAD

WADs Traffic increases Switched On

Indoor WAD

EO Off

EO On

50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 Time [s]

(a) Average Power

(b) Network Power Consumption

Fig. 10. Scenario 4: average power consumption of the WADs when EO is on and off. The network has users increasing their traffic randomly (V = 0) and EO performs load balancing.

Table 4 Comparative of real and ideal energy savings.

Outdoor: Indoor:

Scenario 1

Scenario 2

Scenario 3

Scenario 4

72% (18/25) 96% (32/33.3)

86,62% (20/23.09) 96,45% (31/32.14)

82,29% (19/23.09) 93.34% (30/32.14)

40% (9/22.25) 75.95% (24/31.6)

network traffic is increasing, power save candidates are switched to AM (see black arrows in the Fig. 10b). These have also an effect on the network performance as shown in Table 3. In this scenario, load balancing requirements are different than Scenario 3. In Scenario 3, the users are offloaded to head candidates if their WAD is a power save candidate, and in Scenario 4, the users are offloaded from head candidates back to the power save candidate that are in AM again. Note that in Scenarios 3 and 4, which require handover of clients to other WADs, we do not see a ping-pong effect. This is mainly due to the fact that we cannot turn on and off WADs instantaneously due to (i) the time required for collecting statistics for taking energysaving decisions ( ), (ii) the hardware and software characteristics of the WADs, which influence the switching times (τAM−OM , τOM−AM ), and (iii) the radio wireless technology of the WADs (e.g. WiFi or LTE). Based on these factors, the on-off time intervals may vary from minutes to hours (e.g., when a complete or sector Base Station is considered, which includes components such as air conditioning, amplifier systems, etc.). Nevertheless, methods for detecting and reducing ping-pong effect can be implemented at the CM level for completeness and to avoid QoS degradation in future networks where turning on and off WADs instantaneously may be possible. We also calculate the ideal energy savings based on the network topology assuming a perfect system, where there is no cost for i) switching off and on the WADs (POM = 0 and PAM = 0), and ii) collecting the context information of the users (ECoCo = 0). Additionally, ideal energy savings are calculated assuming the network with and without traffic (V = 0 and V = 0). In case of scenario 1 (V = 0), the ideal energy savings are determined by the ratio of power save candidates to the total number of WADs. In the indoor case, the ideal energy savings are 33.3% (five power save candidates out of a total of 15 WADs). In the outdoor case, these energy savings are up to 25% (one power save candidate out of a total of 4 WADs). In case of Scenarios 2, 3 and 4 (V = 0), the ideal energy savings are determined by the ratio of power consumed by the power save candidates to the power consumed by the total number of WADs (including the power consumed by the generated traffic). Please refer to [34] for checking the model used for calculating the power consumed by WADs receiving and transmitting traffic. In the indoor case, the ideal energy savings are ≈32.14% for Scenarios 2 and 3, and ≈31.6% for Scenario 4. The amount of generated traffic is different in Scenario 4 compared to Scenarios 2 and 3. In the outdoor case, these energy savings are up to ≈23.09% for Scenarios 2 and 3, and ≈22.25% for Scenario 4. The comparison between real and ideal energy savings for the outdoor and

indoor WADs is shown in Table 4. Notice that the power consumed by outdoor WADs is also influenced by climatic conditions, which are difficult to model and predict. The experimental results show how the actions taken for achieving energy savings affect the performance of the network and most importantly, energy savings in practice are close to the ideal savings only for simplified scenarios where traffic and users are static (Scenarios 1, 2 and 3). However, when traffic and network operation become more dynamic like in Scenario 4, the energy savings are reduced. Therefore, an adaptive and context-aware power management framework is necessary to take advantage of different scenarios to save energy as well as maintain minimum effect on the QoS of the users. 6. Conclusions In this paper, we presented results on achievable energy savings in mobile communication networks based on our implementation of an adaptive and context-aware power management framework. Our results quantify energy savings for dynamic scenarios including the energy and time cost of a load-adaptive system. Our framework is targeted to heterogeneous environments, where there is significant device and technology diversity. To this end, we have considered two different types of hardware in our testbed experiments, and confirmed that hardware and software characteristics of WADs play an important role in both deciding which energy-saving mechanism to be used as well as the amount of energy savings that can be achieved. Furthermore, such limitations also affect the granularity with which the adaptations can take place, and hence, the dynamicity of the power-management framework. References [1] Mobile’s green manifesto 2012, (http://www.gsma.com/publicpolicy/wpcontent/uploads/2012/06/Green-Manifesto-2012.pdf). [2] N. Bayer, D. Sivchenko, H.-J. Einsiedler, A. Roos, A. Uzun, S. Göndör, et al., Energy optimisation in heterogeneous multi-RAT networks, in: Proceedings of the IEEE ICIN, Germany, 2011. [3] S. Göndör, A. Uzun, A. Küpper, Towards a dynamic adaption of capacity in mobile telephony networks using context information, in: Proceedings of the ITST, 2011, pp. 606–612. [4] K. Gomez, C. Sengul, N. Bayer, R. Riggio, T. Rasheed, D. Miorandi, MORFEO: saving energy in wireless access infrastructures, in: IEEE WoWMoM, 2013, pp. 1–6. [5] S. Göndör, A. Uzun, N. Bayer, L. Kollecker, A. Küpper, Visualizing the effects of power management algorithms for mobile networks under realistic conditions, in: EGG, IEEE, Berlin, Germany, 2012.

Please cite this article as: N. Bayer et al., Load-adaptive networking for energy-efficient wireless access, Computer Communications (2015), http://dx.doi.org/10.1016/j.comcom.2015.05.004

JID: COMCOM

ARTICLE IN PRESS

[m5G;June 12, 2015;7:19]

N. Bayer et al. / Computer Communications 000 (2015) 1–9 [6] K. Gomez, C. Sengul, N. Bayer, Energy-saving framework for wireless access infrastructures, in: IEEE Online Conference on GreenCom, 2013, pp. 76–77. [7] K. Hooghe, M. Guenach, Toward green copper broadband access networks, IEEE Commun. Mag. 49 (8) (2011) 87–93. [8] A. Fehske, G. Fettweis, J. Malmodin, G. Biczok, The global footprint of mobile communications: The ecological and economic perspective, IEEE Commun. Mag. 49 (8) (2011) 55–62. [9] Cisco Visual Networking Index (VNI): global mobile data traffic forecast, (Update, 2012-2017, White paper, Cisco Systems Inc., 2012). http://www.cisco.com. [10] Potential solutions for energy saving for E-UTRAN, Technical Recommendation 3GPP TR 26.927 V10.0.0, 3GPP, 2011. [11] M. Ajmone Marsan, L. Chiaraviglio, D. Ciullo, M. Meo, Optimal energy savings in cellular access networks, in: Proceedings of the IEEE ICC, 2009. [12] G. Micallef, P. Mogensen, H.O. Scheck, Cell size breathing and possibilities to introduce cell sleep mode, Proceedings of the IEEE EW (2010) 111–115. [13] S.-E. Elayoubi, L. Saker, T. Chahed, Optimal control for base station sleep mode in energy efficient radio access networks, in: Proceedings of the IEEE INFOCOM, 2011, pp. 106–110. [14] A. Chatzipapas, S. Alouf, V. Mancuso, On the minimization of power consumption in base stations using on/off power amplifiers, in: Proceedings of the IEEE GreenCom, 2011, pp. 18–23. [15] A. Conte, A. Feki, L. Chiaraviglio, D. Ciullo, M. Meo, M. Ajmone Marsan, Cell wilting and blossoming for energy efficiency, IEEE Trans. Wireless Commun. 18 (5) (2011) 50–57. [16] M. Lasanen, M. Aubree, C. Cassan, A. Conte, J. David, S.-E. Elayoubi, et al., Environmental friendly mobile radio networks: Approaches of the European OPERA-Net 2 project, in: ICT, 2013, pp. 1–5. [17] G. Dimitrakopoulos, P. Demestichas, W. Koenig, Introduction of cognitive systems in the wireless world: research achievements and future challenges for end-toend efficiency, in: FuNeMS, 2010. [18] M. Ajmone Marsan, C. Guerrero, S. Buzzi, F. Idzikowski, L. Chiaraviglio, M. Meo, et al., TREND: toward real energy-efficient network design, in: SustainIT, 2012. [19] M. Gruber, O. Blume, D. Ferling, D. Zeller, M. Imran, E. Strinati, EARTH: energy aware radio and network technologies, in: PIMRC, 2009, pp. 1–5. [20] L. Saker, S. Elayoubi, H. Scheck, System selection and sleep mode for energy saving in cooperative 2G/3G networks, IEEE VTC-Fall (2009) 1–5.

[21] M. Ajmone Marsan, M. Meo, Energy efficient wireless internet access with cooperative cellular networks, Comput. Netw. 55 (2) (2011) 386–398. [22] K. Dufkova, M. Bjelica, B. Moon, L. Kencl, J.-Y. Le Boudec, Energy savings for cellular network with evaluation of impact on data traffic performance, in: Proc. of IEEE EW, 2010. [23] C. Peng, S.-B. Lee, S. Lu, H. Luo, H. Li, Traffic-driven power saving in operational 3G cellular networks, in: MobiCom ’11, 2011, pp. 121–132. [24] A.P. Jardosh, K. Papagiannaki, E.M. Belding, K.C. Almeroth, G. Iannaccone, B. Vinnakota, Green WLANs: on-demand WLAN infrastructures, Mob. Netw. Appl. 14 (2009). [25] J. Lorincz, A. Capone, M. Bogarelli, Energy savings in wireless access networks through optimized network management, in: Proceedings of the IEEE International Conference on Wireless Pervasive Computing, 2010, pp. 449–454. [26] M. Ajmone Marsan, L. Chiaraviglio, D. Ciullo, M. Meo, A simple analytical model for the energy-efficient activation of access points in dense WLANs, in: Proceedings of the e-Energy, 2010, pp. 159–168. [27] S.-L. Tsao, C.-H. Huang, Review: A survey of energy efficient MAC protocols for IEEE 802.11 WLAN, Comput. Commun. 34 (1) (2011) 54–67. [28] S. Bhaumik, G. Narlikar, S. Chattopadhyay, S. Kanugovi, Breath to stay cool: Adjusting cell sizes to reduce energy consumption, in: Proceedings of the SIGCOMM Workshop on Green Networking, 2010. [29] X. Weng, C. Dongxu, Z. Niu, Energy-efficient cellular network planning under insufficient cell zooming, in: Proceedings of the VTC Spring, IEEE, 2011. [30] J. Lorincz, M. Bogarelli, A. Capone, D. Begusic, Heuristic approach for optimized energy savings in wireless access networks, in: Proceedings of the 2010 International Conference on Software, Telecommunications and Computer Networks (SoftCOM), 2010, pp. 60–65. [31] J. Lorincz, A. Capone, D. Begusic, Heuristic algorithms for optimization of energy consumption in wireless access networks, vol. 5, 2011, pp. 626–648. [32] S. Andreev, P. Gonchukov, N. Himayat, Y. Koucheryavy, A. Turlikov, Energy efficient communications for future broadband cellular networks, Comput. Commun. 35 (14) (2012) 1662–1671. [33] Study on minimization of drive-tests in Next Generation Networks, Technical Recommendation 3GPP TR 36.805 V9.0.0, 3GPP, 2009. [34] K. Gomez, R. Riggio, T. Rasheed, D. Miorandi, F. Granelli, Energino: a hardware and software solution for energy consumption monitoring, Proc. IEEE Winmee 56 (2012) 2506–2521.

Please cite this article as: N. Bayer et al., Load-adaptive networking for energy-efficient wireless access, Computer Communications (2015), http://dx.doi.org/10.1016/j.comcom.2015.05.004

9