Power-Saving Mode for Mobile Computing in Wi-Fi hotspots: Limitations, Enhancements and Open Issues

Transcription

1 Power-Saving Mode for Mobie Computing in Wi-Fi hotspots: Limitations, Enhancements and Open Issues G. Anastasi a, M. Conti b, E. Gregori b, A. Passarea b, Pervasive Computing & Networking Laboratory (PerLab) a Dept. of Information Engineering, University of Pisa Via Diotisavi Pisa, Itay g.anastasi@iet.unipi.it b CNR - IIT Institute Via G. Moruzzi, Pisa, Itay {marco.conti, enrico.gregori, andrea.passarea}@iit.cnr.it Abstract. Nowadays Wi-Fi is the most mature technoogy for wireess-internet access. Despite the arge (and ever increasing) diffusion of Wi-Fi hotspots, energy imitations of mobie devices are sti an issue. To dea with this, the standard incudes a Power-Saving Mode (), but not much attention has been devoted by the research community to understand its performance in depth. We think that this paper contributes to fi the gap. We focus on a typica Wi-Fi hotspot scenario, and assess the dependence of the behavior on severa key parameters such as the packet oss probabiity, the Round Trip Time, the number of users within the hotspot. We show that during traffic bursts is abe to save up to 90% of the energy spent when no energy management is used, and introduces a imited additiona deay. Unfortunatey, in the case of ong inactivity periods between bursts, is not the optima soution for energy management. We thus propose a very simpe Cross-Layer Energy Manager (XEM) that dynamicay tunes its energy-saving strategy depending on the appication behavior and key network parameters. XEM does not require any modification to the appications or to the standard, and can thus be easiy integrated in current Wi-Fi devices. Depending on the network traffic pattern, XEM reduces the energy consumption of an additiona 20 96% with respect to the standard. Keywords: , Wi-Fi, Power-Saving Mode, Network Architecture & Design, Mobie Computing, Network Protocos, Performance of Systems 1 Introduction Since the introduction of the standard in 1997, wireess LANs (aso known as Wi-Fi hotspots) have become more and more popuar. Instaations of Wi-Fi hotspots are nowadays very frequent, for exampe in company and education buidings, coffee shops, airports, and so on. Figure 1 shows a simpe Wi-Fi instaation, where users carrying mobie hosts (e.g., aptops, PDAs,... ) expoit an Access Point to connect to egacy Internet services. This is the scenario used in the paper. Despite its increasing popuarity, Wi-Fi sti presents severa probems that are far to be soved. One of the most important is the energy consumption of wireess interfaces. Wireess cards have shown to account for about 10% of the tota energy consumption in current aptops [1, 7]. This percentage grows up to 50% in handhed devices [1, 31], and even beyond in smaer form-factor prototypes [41]. Even worse, the difference between battery capacities and the requirements of eectronic components is expected to increase in the near future [40]. Energy management is hence a core enabing factor for the Wi-Fi technoogy 1. This work has been carried out whie A. Passarea was with the Department of Information Engineering of the University of Pisa. 1 In this paper we tak about energy management instead of power management, though the atter keyword is quite more diffused in the iterature. Actuay, this paper is not about optimizing the power consumption of , i.e., by adjusting the transmission power or the receiver sensitiveness. Rather, it is about optimizing the time intervas spent by the wireess interface in the different power modes, in order to minimize the energy consumed to perform networking activities. 1

2 INTERNET mobie hosts access point fixed host Figure 1: Wi-Fi hotspot scenario. The standard defines a Power-Saving Mode (), aimed at reducing the energy consumption of mobie devices. Recenty, severa works have been devoted to highight imitations and propose enhancements. The most cosey reated to our work are STPM [1], BSD [30] and S [38] (we provide a comprehensive survey of the reated work in Section 2). These works highight that adds high transfer deays in a range of appication and network configurations. Besides decreasing the user QoS, this might even increase the energy consumption of the device as a whoe, with respect to the case when is not used: the energy saved on the wireess interface by gets overwhemed by the energy spent by the rest of the mobie device during the additiona transfer time. In this paper we focus on a different yet very popuar scenario (thoroughy described in Section 3) where the additiona deay is fairy imited. In Section 4 we show that using in this scenario is an advisabe choice. Thus, in Sections 5 and 6 we extensivey characterize the performance for a wide range of key parameters. To the best of our knowedge, this is the first work providing such a detaied anaysis. We show that is very effective during traffic burst. With respect to the case when it is not used, is abe to save up to 90% of the energy required to downoad a burst. However, is not quite fit to dea with ong User Think Times between bursts, that can actuay represent the main source of energy consumption. From this standpoint, the origina contribution of our work consists in a deep exporation on how to further reduce the energy consumption during User Think Times. Specificay, in Section 7 we define a Cross-Layer Energy Manager (XEM) that expoits information scattered across severa ayers in the protoco stack to detect the beginning of User Think Times and bursts. During bursts XEM activates, whie during User Think Times it switches the wireess interface off. XEM does not degrade the user QoS, and achieves additiona energy saving with respect to the standard between 20% and 90%, depending on the User Think Time ength, and the bursts size. In this work we provide two contributions. In the first part of the paper, we provide an accurate mode of the , and deepy characterize the performance. In comparison with existing works, which usuay highight scenarios in which is not effective, we show that there is a broad range of cases in which can be successfuy used to reduce the energy consumption. In the second part of the paper, we turn to anayze inefficiencies, and propose and evauate XEM. With respect to existing work, XEM smoothy integrates with current and does not require any modification of egacy protocos and appications. XEM is thus a very ightweight, yet efficient, soution to improve in cases in which it is not efficient. 2 Reated work Understanding and enhancing the performance of wireess LANs, mainy in terms of energy saving, has deserved increased attention in the ast few years. Papers in this fied can be divided into two main categories. Some works highight imitations of and propose possibe enhancements. Other works propose energy-management poicies that are not specificay taiored to but can be appied to this technoogy, as we. For ease of reading, in the foowing of this section we foow the above cassification. For the sake of space, and because the environment is 2

3 significanty different, we do not survey the broad research area on energy management for ad hoc networks. 2.1 Energy-management poicies for Wi-Fi hotspots A pioneering work on this topic is presented by Krashinsky and Baakrishnan in [30]. They carry out a simuation anaysis of in presence of Web-browsing traffic. In particuar, they consider a singe mobie user (i.e., no contention) inside the hotspot. The authors of [30] show that can save around 90% of the energy spent by the wireess interface at the cost of highy increased deay in the Web-page downoads. To cope with this probem, they propose the Bounded Sowdown Protoco (BSD). In BSD, the mobie host istens the Access Point Beacons with decreasing frequency during ide times, to be mosty seeping during User Think Times. BSD trades off energy consumption for ower additiona deays. Specificay, if the maximum acceptabe deay is very ow, BSD actuay consumes more energy than. Therefore, BSD can be more or ess suitabe than, if the additiona deay or the energy consumption deserves more importance for the user. As noted in the paper, BSD focuses on a scenario where tremendousy increases the transfer deay, and thus it represents a very effective soution. However, in our scenario this additiona deay is quite imited. It shoud be noted that, thanks to its fexibiity, our Cross-Layer Power Manager (XEM) is abe to use either or BSD during bursts. In contrast to BSD, XEM switches the wireess interface off during User Think Times. As discussed in Section 6.3, the XEM abiity to distinguish interarriva times (for which it is more convenient using the seep state) from User Think Times (for which it is better to switch the wireess interface off) grants greater energy saving. More recenty, Qiao and Shin proposed the Smart Power-Saving Mode (SPMS) [38]. SPMS can be seen as a BSD enhancement. During an ide time, BSD defines staticay the set of points in time where the mobie host istens for Access Point Beacons. Instead, SPMS defines this set of points dynamicay, based on an estimate of the ide time duration. SPMS is more energy efficient than BSD, and sti achieves the same performance in bounding the additiona deay. However, it sti consumes more energy than in some cases, and just expoits the seep state of the wireess interface to conserve energy. Since SPMS is cose to BSD in spirit, the same remarks discussed above appy to SPMS, as we. The authors of [34] propose the Dynamic Beacon Period agorithm (DBP). As BSD and S, DBP aims at reducing the additiona deay introduced by to Web-page downoad times. Basicay, each mobie host seects its own Beacon Interva, and the Access Point is responsibe for generating (custom) Beacon frames for each mobie host. Severa scaabiity issues, that are key points to fairy evauate DBP, are not addressed in [34]. As in the cases of BSD and SPMS, DBP just expoits the seep state of the wireess interface to conserve energy, for any kind of ide time that might occur. Anand et a., [1] carry out an experimenta evauation of both on PDAs and aptops. They primariy focus on the traffic generated by appications using network fie systems such as NFS and Coda. Their resuts confirm the concusions in [30], as far as the additiona deay introduced by the. To overcome this probem, they propose the Sef-Tuning Power Management (STPM) protoco. STPM operates at the Operating System eve, and expoits hints provided by the network appications. Essentiay, hints describe the near future requirements of appications in terms of networking activities. STPM expoits these hints, and the energy characteristics of the entire system, to manage the wireess interface appropriatey. When these hints are not avaiabe, STPM estimates the traffic patterns by spoofing it. Like STPM, our Cross-Layer Energy Manager sits on top of different energy management poicies, and dynamicay chooses the most appropriate one. The main difference between [1] and our work is that XEM is simper, and never requires coaboration from the appications, i.e., no modifications of the appication code is required. Again, [1] focuses on a scenario where deays are a big probem, whie in our scenario they are not. Finay, [9, 39] propose energy-management poicies for WLAN that are orthogona to the work presented in this paper, and hence can coexist with XEM. 3

4 2.2 Energy-management poicies for generic wireess LANs Other works face the energy-management probem in WLAN environments, but do not focus on a specific wireess technoogy. The authors of [32] propose a soution entirey centraized at the Access Point. Time is divided in Beacon Intervas (as in the standard ), and at the beginning of each Beacon Interva the Access Point computes a schedue for transmitting frames during the coming Beacon Interva. Before any other transmission, the Access Point broadcasts a Beacon Frame to pubicize which mobie hosts are going to receive frames. These mobie hosts remain awake unti they have received a the schedued frames, whie the other hosts can immediatey switch to a ow-power mode. This soution gives to the Access Point the fexibiity of impementing severa scheduing poicies, but requires i) significant computationa burden at the Access Point, and ii) non-trivia modifications to the standard. The authors of [6] design an energy manager taiored excusivey to Web-based appications. By means of prefetch-ike techniques, Web pages are transferred over the WLAN in a singe (or few) burst, thus maximizing the amount of time during which the wireess interface is switched off. This technique does not introduce significant additiona deays. Of course, the energy manager is tied with the particuar appication it is designed for. In [3] it is shown that this constraint can be reaxed with an acceptabe degradation of the energetic performance. Specificay, [3] dynamicay estimates the expected duration of ide times. The mobie host is switched off for the (predicted) duration of the ide time. The work in [6] and [3] inspired some ideas on how User Think Times and new bursts can be detected. However, XEM fuy expoits when appropriate, and can avoid reying on appication-eve information. The works in [31, 44, 46] use inactivity timeouts to decide when to switch off the wireess interface. Timeout vaues are fixed, and depend on the specific appication. [31] reies on an Indirect-TCP architecture and buffers at the Access Point packets arriving whie the mobie host is disconnected. Instead, [44] avoids any support from the Access Point, and expoits knowedge of the appication behavior to avoid missing packets. Aso [46] uses a pure cient-centric approach, i.e., no support from the Access Point is expoited. Specificay, [46] uses an approach very simiar to [3], in the sense that interarriva times are estimated on-ine. Furthermore, inactivity timeouts are used to detect User Think Times. With respect to [3], no support from the Access Point is expoited. Hence, packets that may arrive whie the mobie host is disconnected are ost. Inactivity timeouts are aso used by XEM. However, in our system they are dynamicay adjusted based on the status of the network path. The works in [33, 17, 42] advocate energy management at the operating system eve. [33] expoits on-ine appication-eve hints to decide when to shut down the wireess network. Hence, this system requires modifications to the appication code. The authors of [17, 42] formuate the energy-management probem as a inear program, where the objective is minimizing the energy consumption of a particuar component, and the maximum toerabe performance degradation (for exampe in terms of additiona deay) is the constraint. Then, they derive optima energy management poicies to drive the component in the different operating modes. The main drawback of this approach is that it requires a-priori statistica modes of the component usage. This information is not required by XEM. Finay, other approaches to energy management incude transmission power contro techniques [22], or a drastic re-design of the appication-eve architecture [37, 24, 25]. Specificay, [25] introduces a quite recent technoogy, named AJAX. The main idea is decouping (in Web-ike appications) the user and the server via a proxy-based component (caed Ajax engine) running on the cient. The user actuay interacts with the Ajax engine, that asynchronousy fetch from the server the data required to fufi the user requests. Ajax is abe to significanty reduce the amount of data exchanged over the network in case of sight modification of a currenty-rendered Web page. In genera, appication-eve techniques are orthogona to XEM. In the case of AJAX, the most straightforward interaction we can see is using XEM beow AJAX. XEM woud interpret the traffic pattern generated by AJAX (instead of that generated by the user), and manage the wireess interface accordingy. 4

5 downoad interva T fixed host t INTERNET mobie hosts access point fixed host tagged mobie host bursts d k User Think Times UTT k number of bursts = number of UTTs = N BR t (a) Wi-Fi hotspot scenario (b) Appication-eve traffic Figure 2: The reference environment. 3 Detaied Scenario and Evauation Methodoogy 3.1 Reference Scenario In our anaysis, we consider the typica Wi-Fi hotspot scenario, depicted in Figure 1 and repicated in Figure 2(a) for the reader convenience, in which a mobie user accesses the Internet through an Access Point. We focus on best-effort Internet appications, such as Web browsing, e-mai, fie transfer (hereafter referred to as reference appications). This choice is motivated by the evidence that the traffic generated by these appications represents the ion s share of the today Internet traffic, and they are very ikey to be the dominant appications aso in the near-future Internet [10]. Figure 2(b) shows a snapshot of the typica traffic generated by the reference appications. A tagged mobie host downoads a predefined number of bursts (N BR ) from a fixed server connected to the Internet. The downoad of two consecutive bursts is separated by a User Think Time (UTT) during which no traffic fows between the server and the mobie host (the other detais shown in the Figure are reated to the mode, and wi be thus expained in Section 5). Though very simpe, this traffic mode captures the typica user behavior for non rea-time appications. For exampe, Web users downoad a page (i.e., a burst) and then read the page contents without generating any traffic on the network. By considering severa such downoads (say, N BR ) from the same site, we mode the behavior of a user navigating a singe Web site for a whie. Because of its importance, some concepts of the paper are presented by using the Web as the reference appication. However, the traffic mode is genera enough to represent aso other best-effort appications, as we. Furthermore, we investigated a broad range of the scenario parameters when evauating and XEM. Therefore, we beieve our resuts are not vaid ony in the Web case. Furthermore, as far as the XEM definition, it is not heaviy tied to the Web case, and can work with different appications, as discussed in Section 7.4. We assume that the mobie host communicates with the fixed server through a standard TCP-Reno (without deayed acks [45]) connection. We aso assume that consecutive bursts be downoaded over the same connection. In the Web case this corresponds to using the persistent-connection option defined by HTTP/1.1 [27]. Since HTTP fie transfers usuay consist of few KB [13, 20, 21] this option was defined to avoid the huge overhead of opening a new TCP connection for each fie transfer. It reduces downoad times, and aows TCP to precisey earn the path congestion. In Section 4 we highight that this option has further advantages when is used. Specificay, the additiona deay introduced by to TCP transfers becomes fairy sma, and it does not significanty impact on the energy consumption of the device as a whoe. Thus, using persistent connections is a good idea for the other reference appications, as we. We finay assume that the mobie host does not utiize parae concurrent connections to downoad bursts. This is aigned with the suggestions of [27] when persistent connections are used, 5

6 and it makes the anaysis of both and XEM simper. In Section 7.4 we highight how XEM can be extended to work in the case of concurrent TCP connections. One might argue that anyway the egacy TCP/IP architecture exhibits poor performance in a WLAN environment, both in terms of throughput and energy consumption [12]. However, TCP/IP is currenty the ony off-the-shef soution for Wi-Fi hotspots and thus our environment is simiar to rea-word WLAN instaations. In our scenario, the hotspot is popuated by other N (background) mobie hosts in addition to the tagged mobie host. We assume that, at each point in time, M mobie hosts out of N are active, i.e., they have a frame ready to be sent. As discussed in Section 6.2.3, by varying the number of active mobie hosts (i.e., M) we can anayze the sensitiveness of to the contention eve in the hotspot, and therefore its scaabiity with respect to the number of users sharing the same Access Point Power-Saving Mode () As a significant part of this work is devoted to anayze the performance, in this section we briefy reca the main features of this agorithm. The interested reader is referred to the IEEE standard for a compete description [29]. The objective of the is to et the wireess interface of a mobie host in the active mode ony for the time necessary to exchange data, and turn it in seep mode whenever it becomes ide. In a Wi-Fi hotspot, this is achieved by expoiting the centra roe of the Access Point. Each mobie host within the hotspot ets the Access Point know whether it utiizes the or not. Since the Access Point reays every frame from/to any mobie host, it buffers the frames addressed to mobie hosts using the Power-Saving Mode. Every Beacon Interva usuay, 100 ms, the Access Point broadcasts a specia frame, named Beacon (Figure 3(a)). This frame contains a Traffic Indication Map (TIM) that indicates mobie hosts having at east one frame buffered at the Access Point. mobie hosts are synchronized with the Access Point, and wake up to receive Beacons. If they are indicated in the TIM, they downoad the frames as is shown in Figure 3(b). Specificay, the mobie host sends a specia frame (ps-po) to the Access Point by means of the standard DCF procedure. Upon receiving a ps-po, the Access Point sends the first data frame to the mobie host, and receives the corresponding ack frame. If appropriate, the Access Point sets the More Data bit in the data frame, to announce other frames to the same mobie host. To downoad the next frame, the mobie host sends another ps-po. When, eventuay, the mobie host has downoaded a the buffered frames, it switches to the seep mode. To send a data frame, a mobie host (if the case) wakes up and performs the standard DCF procedure. Specificay, the mobie host sends the data frame, and receives an ack frame from the Access Point (Figure 3(c)). To summarize, a mobie device operating in Power-Saving Mode is required to be awake to perform three basic operations: (i) receiving Beacon frames; (ii) downoading data frames from the Access Point; and (iii) sending data frames to the Access Point. This remark is fundamenta for the anaytica characterization of that have been derived in [4] presented in Section Evauation Methodoogy In the environment described above, one of the main inefficiencies in energy usage is istening during ide times. It is we known that the traffic generated by the reference appications exhibits different types of ide times [21, 20, 13, 8]. Specificay, ide times inside traffic bursts (referred to as interarriva times) are typicay very short, ess than 1 s [21, 20, 13]. On the other hand, ide times between consecutive bursts (referred to as User Think Times), are onger and may ast up to 60 s and beyond [21]. As the goa of is reducing the energy consumption during ide times, we extensivey anayze its behavior with respect to both interarriva times and User Think Times. 6

7 SIFS AP tagged mobie host tb PIFS Beacon Interva Beacon PIFS Beacon t t AP tagged mobie host PIFS Beacon DIFS tmac Beacon Interva ts PS Po SIFS Data SIFS ACK PIFS Beacon t t AP tagged mobie host DIFS tmac ta Data ACK t t active active active seep seep seep active (a) receiving beacons DCF procedure (b) receiving frames DCF procedure (c) sending frames Figure 3: operations. The amount and duration of interarriva times is dictated by both the appication-eve protocos, and the characteristics of the network path between the mobie and the fixed host. Thus, we anayze the performance with respect to key appications and network parameters, i.e., i) the average burst size; ii) the transport-eve throughput; and iii) the MAC-eve contention (i.e., the number of users in the same Wi-Fi hotspot). We then study the performance during User Think Times. We show that the energy consumption during these phases may dominate the energy consumption due to the whoe traffic pattern. Since is far from optima during UTTs, we define and evauate a Cross-Layer Energy Manager (XEM) that drasticay reduces energy consumption during UTTs. Our anaysis reies on both anaytica and simuation resuts. Specificay, we extend the simuation mode used in [16] to impement the reference network scenario described above (see Section 6 for detais). Furthermore, to better understand the behavior shown by simuation, we expoit an anaytica mode 2 we derived in [2, 4, 36], that provides cosed formuas for the energy consumption in the cases where is used or not. A brief presentation of this mode is given in Section Performance Indices Our anaysis is mainy based on the foowing performance figures: E C : the average amount of energy spent by the wireess interface to downoad N BR bursts from the fixed to the tagged mobie host, when is not active (i.e., in continuous active mode, CAM); E P : the average amount of energy spent by the wireess interface to downoad N BR bursts from the fixed to the tagged mobie host, when is active; R (E P, E C ): the ratio between the above indexes. This is a key index, since it shows the fraction of energy spent when is active, with respect to the case when no energy management is used, and, thus, it shows the efficiency. Throughout the paper, we anayze energy consumption breakdowns for the different energy-saving poicies under investigation. In that cases, more specific performance indexes are defined. When meaningfu, the index R (, ) is aso appied to coupes of those indexes, to show the reative advantage of the first one with respect to the second one. Admittedy, most of our anaysis negects the energy consumption of mobie-device components other than the wireess interface. Resuts discussed in the next section show that in our scenario this is a reasonabe choice. 2 As shown in [2, 4, 36], anaytica and simuation resuts fuy agree. 7

8 % 50% 45% 10% 9% 0 this work zone 10% impact on the overa system 15% f f idea energy saving due to wireess interface energy management ony f 1+f 90% PDA LAPTOP PDA d % id % 100% 90% 75% STPM idea BSD idea S idea in our scenario Energy saving due to wireess interface management ony f theoretica imit f= LAPTOP STPM, BSD, S zone 20% 0 LAPTOP 15% PDA d % (a) (b) Figure 4: impact on the overa system (a), and on the wireess interface ony (b). 4 Effects of deay on energy consumption Thanks to the resuts in [1, 30, 38] it is now we understood that the additiona deay introduced by to TCP transfer times may even make counterproductive from an energy-saving standpoint. Even though much energy spent on the wireess interface can be saved by, the other device components continue to drain energy during the additiona transfer time. This cost may competey overwhem the energy saved on the wireess interface. For exampe, Anand et a. [1] measure sowdown factors as high as 16x to 32x when using. It can be noted that high additiona deays arise when short TCP connections are used over short RTT paths. Indeed, [1, 30, 38] measure high additiona deays in cases where the RTT between the mobie and fixed hosts (measured when is not active) is very short (few tens of miiseconds). Moreover, they focus on Web traffic without persistent connections [30], and NFS-ike traffic [1] 3, which actuay generate very short TCP connections. Due to the TCP 3-way handshake, and to the sow-start agorithm, a new TCP connection requires severa RTTs even to fetch a few KBytes. Furthermore, the work in [30] shows that rounds every RTT up to the next 100ms, due to the beaconing mechanism. This can be a very high additiona deay for short RTTs (i.e., in the order of few tens of miiseconds). Therefore, when is used to downoad data over short-ived, non-persistent TCP connections, in case of short RTTs, the additiona deay can be very high. We purposey choose a different scenario for our anaysis. As mentioned in Section 3, we consider persistent TCP connections, as suggested by HTTP/1.1 [27]. Furthermore, we focus on a broader RTT range (measured when is not active), in the order of few hundreds of miiseconds. Even though Web proxies and caches tend to reduce the RTT, they cannot be used with any Web content (e.g., cannot be used with dynamicay generated pages). Moreover, it is sti common to measure RTTs in the order of 200 ms and above whie accessing popuar servers over intercontinenta paths (e.g., accessing ebay.com, cnn.com, nasdaq.com, amazon.com between Europe and US). In this case, fetching data requires ess RTTs, because the congestion window is aready stabe, since the connection is persistent. Furthermore, the cost of rounding up every RTT to the next 100ms is reduced if the origina RTT (measured without ) is aready a few hundreds of miiseconds. Indeed, in our scenario the additiona deay that we have measured, averaged over a the experiments presented in the foowing, is just around 15% of the origina transfer time (corresponding to a 1.15x sowdown). To understand the impact of this sowdown on energy consumption, we foow a simpe anaytica approach. When is not used, we assume that both the wireess interface and the rest of the system constanty drain a 3 Anand et a. in [1] focus on other types of traffic as we. However, a the traffic patterns for which introduces high deays share the same features. 8

9 fixed amount of power, denoted as P (N) and P (B), respectivey. Let us denote by t the downoad time of a burst, and by e (N) C and e (B) C the energy spent in continuous active mode during t by the wireess interface and the rest of the system, respectivey. Let us finay denote by f the reative cost of the wireess interface with respect to the rest of the system, i.e. f = P (N) /P (B). Thus, the energy spent in the burst downoad is e C = e (N) C + e(b) C = P (N) t + P (B) t = (1 + f)e (B) C. The use of has two effects. On the one hand, it reduces e(n) C. Let us denote this energy saving by β, i.e., e (N) P = βe (N) C where e(n) P is the energy spent on the wireess interface when is used. On the other hand, increases the energy of the rest of the mobie host because of the additiona deay. If d denotes this additiona deay as a fraction of t, then we obtain e (B) P = (1 + d)tp (B) = (1 + d)e (B) C. It is now easy to evauate the overa energetic advantage brought by. Specificay, we define the index as = (e C e P )/e C, where e P = e (N) P + e (B) fβ+1+d P. After simpe manipuations we obtain = 1 1+f. Figure 4(a) pots as a function of d for various vaues of f. Typicay, f increases as the device form-factor shrinks; representative vaues for a aptop and a PDA are 1/9 and 1, respectivey [1]. Characterizing β is the task of most part of this paper. However, we can here anticipate that β = 0.1 is a reasonabe vaue to have a first rough yet significant picture. This vaue aso matches other resuts in the iterature [30, 38]. Figure 4(a) ceary differentiates our work from [1, 30, 38]. STPM [1], BSD [30] and SPMS [38] are mainy designed to operate in cases when the additiona deay is arge (e.g., the 16x sowdown measured by [1] corresponds to d = 1500%!). Indeed, in these cases drops beow 0, stating that actuay produces an energy increase on the whoe device. Instead, our work focuses on a region where d is imited, and becomes effective, mosty for sma form-factor devices. For this cass of devices, saves a arge portion of the energy consumption due to networking activities, without charging significanty the other device components. Based on these resuts, hereafter we measure the energy consumption of the wireess interface. Figure 4(a) shows the theoretica imits achieved by an idea poicy that competey eiminates the wireess interface energy consumption (this poicy ceary represents the asymptotica imit of any energy management technique focused on the wireess interface). On a burst downoad asting t seconds, the idea poicy consumes e I = e (B) C. Figure 4(b) compares more thoroughy the performance with e I. Specificay, it pots the index id, which is defined as the ratio between the energy saved by, and the energy saved by the idea poicy, i.e., id = (e C e P )/(e C e I ) = 1 β d f. In our scenario, achieves 75% of the idea energy saving. It is interesting to note that the additiona deay reduces the energy saving just by 15%. Furthermore, the id index can be used aso to roughy understand the maximum expected improvement of STPM, BSD and SPMS over in our scenario. STPM activates or deactivates based on predictions about the future traffic profie. This way, it reduces the additiona deay to negigibe vaues. In the best possibe case, it spends on the wireess interface the same energy spent by, without increasing the energy consumption of the rest of the device. Setting d = 0 in the id formua thus gives the maximum energy saving of STPM. Getting anaytica resuts for BSD and S is not straightforward. However, by inspecting the resuts provided in [30, 38] we can sti derive some imit. S is generay abe to avoid the additiona deay, and in severa cases achieves the same wireess interface energy consumption of. Thus, the id vaue for d = 0 is a good indication for the maximum S performance, as we. Aso BSD is abe to reduce the additiona deay to negigibe vaues. [30] shows that this is often achieved without increasing the wireess interface energy consumption with respect to. We thus consider the same maximum energy saving for BSD, as we. These optima vaues are indicated in Figure 4. STPM, BSD, and S seem abe to improve the performance of aso in our scenario. Nevertheess, we beieve that sti represents a vaid option, because i) it achieves significant energy saving anyway, ii) it is aready avaiabe on most of the commercia devices, and iii) it is thus a free-of-charge soution. It shoud aso be noted that the rea performance of STPM, BSD and S can be ower than the vaues in Figure 4(b). As an exampe, [30, 38] show that, in order to eiminate the additiona deay, BSD might increase the energy spent on the wireess interface with respect to. On the other hand, in order to keep the same energy saving, BSD must introduce additiona deays around 14%. Simiar remarks suggest that, 9

10 though being very effective when d is high, BSD, STPM and S do not perform far from in our scenario. As a fina remark, it shoud be noted that the above discussion appies to the burst downoad phases. We postpone a simiar discussion about UTTs to Section 7, to have the chance of incuding aso XEM in the picture. The above remarks show that there is a broad range of cases in which using as an energy-saving technique is advisabe. Therefore, we now anayze in depth the performance in terms of energy saving. 5 Anaytica mode With reference to the network scenario and the appication-eve traffic mode depicted in Figure 2(a) and Figure 2(b), respectivey, in this section we derive a mode for evauating the average energy spent by the tagged mobie host to downoad N BR bursts from the fixed host (any two consecutive bursts are separated by a User Think Time). Due to space reasons, we here present the main anaytica resuts, and we skip many detaied proofs. Interested readers can refer to [2, 4, 36] for a the detais. To mode the tagged mobie-host behavior we utiize the foowing approach. We repicate n times the downoad of N BR bursts, and focus on the generic i-th repica. E (i) P and E (i) C denote the energy spent during the i-th repica when is enabed and disabed, respectivey. In the foowing, we derive cosed formuas for E (i) and E (i) C, and { show that E (i) P E C as foows: }i,...,n and { E (i) C By introducing the cosed formuas for E (i) P and E C. } i,...,n are composed by i.d. random variabes4. Therefore, we can express E P and n i=1 E P = im E(i) P n n E C = im n and E (i) C n i=1 E(i) C n in Expressions 1, we finay obtain the cosed formuas for E P As far as E (i) C, it is worth noting that, when is disabed, the wireess interface of the tagged mobie host is aways active. Hence, if T (i) denotes the duration of the i-th repica (aso referred to as the downoad interva), and P ac denotes the power drained by the tagged mobie host in the active mode, E (i) C can be expressed as E (i) C = T (i) P ac. (2) On the other hand, when is enabed, the tagged mobie host remains active just for a portion (T ac (i) ) of the downoad interva 5, whie it is seeping for the rest of the time (T (i) s ). Therefore, if P s is the power drained by the tagged mobie host in the seep mode, E (i) P can be expressed as foows: E (i) P Equations 2 and 3 show that both E (i) P T (i) and T ac (i), respectivey. = T (i) ac P ac + T (i) s P s = T (i) ac (P ac P s ) + T (i) P s. (3) and E (i) C depend on T (i) and T (i) ac. In the foowing subsections we derive P (1) 5.1 Modeing the downoad interva (T) With reference to a generic i-th repica, T (i) may be thought of as made up of two components (see Figure 2(b)): i) the tota time during which bursts are downoaded (T (i) data ), and ii) the tota inactive time due to User Think Times (T (i) ide ). Denoting by td(i) k the time required by the tagged mobie host to downoad the k-th burst in the i-th repica, and by UTT (i) k the duration of the k-th User Think Time in the i-th repica, T (i) can be written as foows: n 4 To simpify the notation, in the foowing we omit indicating the range of variabiity of i, e.g. 5 T (i) ac aso incudes the transition times from the seep to the active mode. E (i) P oi,...,n is referred to as n E (i) P o 10

11 T (i) = T (i) data + T (i) ide = N (i) BR k=1 td (i) k + N (i) BR k=1 UTT (i) k. (4) { } { } It can be shown that T (i) data and T (i) ide are composed by identicay distributed random variabes. Furthermore, for each coupe i, k, N (i) BR and td(i) k, as we as N(i) BR and UTT(i) k, are mutuay independent. It is aso worth pointing out that we assume that TCP aways works in the steady state (i.e., we do not consider sow-start phases), which is a common assumption in the iterature [35], and very reasonabe in the case of persistent connections. Furthermore, to simpify the anaysis, we approximate the steady-state TCP throughput with a constant vaue 6, hereafter referred to as γ TCP. Therefore, if E [d] denotes the average burst size, after simpe manipuation the average vaue of the downoad interva can be expressed as: E [T] = E [N BR ] { } E [d] + E [UTT] γ T CP 5.2 Modeing the time spent in the active mode (T ac ). (5) Since we are assuming a TCP/IP architecture, the traffic on the WLAN reated to the tagged mobie host incudes: i) TCP segments 7 coming from the fixed server; ii) TCP acks sent by the tagged mobie host to the fixed server; and iii) Beacon frames periodicay broadcast by the Access Point. Thus, T (i) ac is the time spent in the active mode by the tagged mobie host to hande these traffic components. Before proceeding on, we need to introduce some assumptions and emphasize some properties reated to our mode. Property 1. In the foowing, we assume that Beacon frames are safey transmitted, i.e., they do not coide with transmissions from any other mobie host. In other words, the MAC protoco guarantees that the shared medium is ide at the beginning of each Beacon Interva, and no transmissions are attempted unti the Beacon frame is received (see Figure 3(a)). This assumption is aigned with the most up-to-date proposas within the working groups [28]. Property 2. Let us define a sequence of frames as a set of frames exchanged between the mobie host and the Access Point, where each frame is spaced from the previous one by a SIFS interva. According to the DCF definition [29], in our WLAN environment ony the first frame of a sequence can undergo coision. In other words, either the first frame of a sequence coides, or the whoe sequence is safe. Property 3. Each TCP segment sent by the fixed host to the tagged mobie host is encapsuated into a distinct IP packet. Thus, if we assume that both IP- and MAC-eve fragmentation are disabed, the tagged mobie host downoads each TCP segment from the Access Point inside a distinct data frame. Since we aso assume that the RTS/CTS mechanism is disabed, downoads occur by exchanging a sequence of frames incuding a ps-po, data, and ack frame (between the tagged mobie host and the Access Point), as shown in Figure 3(b). Simiary, each TCP ack is upoaded to the Access Point inside a distinct data frame, i.e., by exchanging a sequence of frames composed by a data and an ack frame (see Figure 3(c)). Property 4. Let i) s (i) j be the time required by the tagged mobie host to downoad the generic j-th TCP segment during the i-th repica, starting from the point in time when the tagged mobie host starts the DCF procedure to send the reated ps-po frame; ii) a (i) r be the interva required by the tagged mobie host to upoad the r-th TCP ack during the i-th repica, starting from the point in time where the tagged mobie host starts the DCF procedure to send the reated data frame; and iii) b (i) be the time required by the tagged mobie host to receive the -th Beacon frame during the i-th repica, starting from the beginning of the reated Beacon Interva. Then, for any tripe j, r,, 6 The vaidation of the anaytica mode carried out in [2, 4, 36] shows that these assumptions do not compromise the accuracy of the anaytica resuts. 7 For the sake of simpicity, we indicate TCP segments containing appication data as TCP segments, whie TCP acks denote TCP segments containing just acknowedgments. 11

12 it can be shown that the time intervas s (i) j, a(i) r and b (i) do not overap. Based on the above properties, T ac (i) can be regarded as the sum of times required to i) receive a the Beacon frames, ii) downoad a the TCP segments, and iii) upoad a the TCP acks, within a downoad interva i.e. N T ac (i) = (i) seg j=1 s (i) j + N (i) ack r=1 N b (i) a (i) r + =1 b (i). (6) In (6), N seg, (i) N (i) ack and N(i) b are the number of TCP segments, TCP acks, and Beacon frames exchanged (between the mobie host and the Access Point) during the i-th repica, respectivey. In our mode: i) the number of TCP segments downoaded is equa to the number of TCP acks upoaded (i.e., N seg (i) = N (i) ack ); and ii) b(i) can be reasonaby approximated with a constant vaue, throughout referred to as b. Finay, by anayzing the properties of the random variabes N seg, (i) N (i) ack, s(i) j and a (i) r, it can be shown that the average vaue of T ac can be expressed in a very intuitive way: E [T ac ] = E [N seg ] (E [s] + E [a]) + E [N b ] b. (7) Deriving cosed formuas for E [s] and E [a] woud require a detaied anaysis of the DCF function. The compete mode accounting for a the DCF detais (retransmissions, contentions, etc.) is derived in [2, 4, 36], and is here omitted for the sake of space. The main issue to be highighted here is that both E [s] and E [a] incude two components, i.e. the average MAC deay and the average sequence time (see Figure 3(b,c)). The average MAC deay is defined as the interva between the time when the DCF procedure is invoked to transmit a frame, and the time of the successfu transmission (possiby after a number of unsuccessfu attempts). Intuitivey, this component is statisticay equivaent for both TCP segments and TCP acks, and corresponds to the time spent by the tagged mobie host in the DCF procedure to send the ps-po frame and the data frame containing the TCP ack, respectivey (see Figure 3(b,c)). The average sequence time is defined as the interva between the time when the transmission of the first frame in a sequence starts, and the time when the reception of the ast frame in that sequence ends. Ceary, the average sequence time depends on the frames in the sequence, and is thus different for TCP segments and TCP acks (see Figure 3(b,c)). The ast step to derive a cosed formua for E [T ac ] is evauating E [N seg ] and E [N b ]. It can be shown that the average number of TCP segments exchanged during a downoad interva (E [N seg ]) is equa to the average size of a bursts downoaded in that repica, divided by the Maximum Segment Size (MSS) of the TCP connection, i.e., E [N seg ] = E [N BR] E [d] MSS. (8) In addition, the average number of Beacon frames received by the tagged mobie host during a repica (E [N b ]) is the ratio between the average downoad-interva duration, E [T], and the duration of a Beacon Interva, BI: E [N b ] = E [T] BI. (9) By substituting Equations 8 and 9 into Equation 7 we obtain the foowing cosed formua for E [T ac ]: E [T ac ] = E [N BR] E [d] MSS (E [s] + E [a]) + E [T] BI b. (10) Finay, by introducing Equations 2 and 3 into Expression 1, after simpe agebraic manipuations, E C and E P can be expressed as foows: { where E [T] and E [T ac ] are given by Equations 5 and 10, respectivey. E C = E [T] P ac E P = E [T ac ] (P ac P s ) + E [T] P s, (11) 12

13 Energy in CAM Energy in Energy(J) anaysis simuation active mobie hosts (M) Energy (J) anaysis simuation active mobie hosts (M) (a) CAM (b) Figure 5: Exampe of vaidation pots. 6 Evauating the Power Saving Mode As mentioned in Section 3, the performance anaysis of is carried out by using both simuation and the anaytica mode derived in the previous section. Specificay, anaytica resuts are used to provide better insights in the behavior highighted by simuation. As an exampe of the agreement between the anaytica and the simuation mode, Figure 5 shows two of the vaidation pots for E C and E P presented in [36]. According to the ide-time cassification presented in Section 3, we anayze performance by considering separatey its behavior during interarriva times i.e., during bursts (Section 6.2), and during User Think Times (Section 6.3). 6.1 Simuation Environment Our simuator extends the mode used in [16], and impements the reference environment described in Section 3. It simuates a fu-compiant hotspot (popuated by a variabe number of background mobie hosts), and fu-compiant TCP-Reno between the mobie and the fixed host. Pease note that the simuation mode impements a the features of both and TCP. To aow for significant vaues of burst sizes and User Think Times we make reference to the Web traffic. Therefore, each burst corresponds to the downoad of a Web page. In particuar, we consider the statistica modes of the Web traffic presented in the we-known works by Crovea et a. [13, 20]. A typica simuation run proceeds as foows (Tabe 1 summarizes the defaut vaues for the main simuation parameters). The tagged mobie host downoads N BR bursts from the fixed server (reca that two consecutive bursts are spaced by a User Think Time). The average Web-page size and User Think Time duration (i.e., E [d] and E [UTT] in Tabe 1) are derived from [13, 20]. However, we tested the system over a wide range of burst and UTT vaues, making the anaysis vaid for the genera traffic mode presented in Section 3, and not ony for the Web case. To mimic a reaistic TCP connection between the mobie host and the fixed server, Internet Round Trip Times (as woud be measured without ) are samped from an exponentia distribution (the defaut average vaue RTT is reported in Tabe 1). To simuate packet osses at Internet routers, TCP segments are randomy dropped with probabiity p tcp. Note that p tcp just accounts for osses in the wired network, due to routers buffer overfow. The additiona packet oss due to the WLAN depends on the MAC protoco behavior, and is thus not a simuation parameter (it can actuay be derived by simuation). Finay, energy parameters are as foows. The power consumptions in the seep and active modes (i.e., P s and P ac ) are the same as those used in [30]. These 13

14 Parameter Vaue Unit Parameter Vaue Unit N BR MSS 1460 B E [d] KB P s 50 mw E [UTT] 3.25 s P ac 750 mw RTT 150 ms t sa 1 ms p tcp 1% - BI 100 ms Tabe 1: Defaut simuation parameters are quite simiar to vaues used in other we-known anayses [23], and comparabe to recent datasheets [18]. t sa denotes the time required by the wireess interface to switch from the seep to the active mode. Note that this parameter aows us to aso incude the cost of switching between the wireess interface operating modes according to. Its defaut vaue is derived from the measurements in [30]. Specificay, [30] measured that, whie operating in mode, the wireess interface spends about 2 ms to switch from the seep to the active mode and to receive a Beacon Frame. Based on the standard [29] it is easy to show that the time required to receive a Beacon Frame is about 1 ms. Therefore, we assume 1 ms as the time required by the hardware to switch from the seep to the active mode. We do not consider the impact of different t sa vaues on, because we focus more on networking and appication-eve parameters. Note that the switching time between operating modes (and thus the reated costs) coud be reduced significanty by improved hardware design (simiary to what is happening for channe switching in the mesh networks domain). Finay, the vaue of the Beacon Interva (BI) is the one suggested by the standard [29]. To increase the resuts reiabiity, each simuation experiment is repicated 10 times. Confidence intervas reported throughout the paper have 95% confidence eve. 6.2 performance during bursts Bursts and interarriva times are determined by both appication and networking protocos. In our scenario, where data mainy fow from the fixed server to the tagged mobie host, the appication dictates the burst sizes 8, whie the TCP protoco is the main responsibe for interarriva times. Thus, we now focus on the impact of two parameters, i.e. i) the average burst size (Section 6.2.1), and ii) the TCP-connection throughput (Section 6.2.2). In both cases, we assume a singe mobie host in the hotspot (i.e, M = 0). Section extends the anaysis by considering severa mobie hosts in the same hotspot (i.e., M > 0) Impact of the burst size As mentioned above, in our mode each burst corresponds to the entire downoad of a Web page. There is a wide consensus about the type of distribution for modeing page sizes (see, for exampe, [13, 20, 8]). On the other hand, the average vaue of this distribution can be highy variabe, and can range from 20 KB up to few MB [13, 20, 21]. Based on these remarks, in our simuation mode the burst-size distribution is defined by the random variabe a S, where: i) a is a (integer) scaing factor, and ii) S is the random variabe defining the page size distribution derived in [13, 20]. The average burst size can thus be scaed (by varying a) without modifying the distribution s coefficient of variation. This aows us to evauate under reaistic traffic oads. Specificay, we report a set of experiments where a varies between 1 and 100, whie the average of S (denoted by µ) is set to 20 KB [13, 20]. This resuts in an average burst size ranging from 20 KB to about 2 MB. We beieve this range aso represents the traffic when techniques such as oss-ess compression or AJAX [25] are used, i.e. techniques that significanty reduce the amount 8 E.g., in the Web case the burst sizes are determined by the content the user is downoading. 14

15 Energy (J) CAM scaing factor (a) R(Ep,Ec) scaing factor (a) (a) Energy pots (b) R (E P,E C) Figure 6: performance as function of the average burst size (a µ). of data exchanged over the network per user request. More in genera, we beieve that this range represents a our reference appications. As in this set of experiments we intend to investigate the performance during bursts, User Think Times are aways set to 0. Since the TCP-connection evoution depends on i) the average Round Trip Time (RTT) and ii) the segment-oss probabiity (p tcp ) [35], and both parameters can be reasonaby assumed to be independent of the User Think Time duration, setting UTT to 0 is justified. Figure 6(a) pots E P (bottom curve) and E C (top curve) for different average burst sizes. The most interesting feature is that energy increases ineary in both cases. This behavior can be expained by means of Equation 11. Since we assume E [UTT] = 0 and E [d] = a µ, E C becomes: E C = P ac E [N BR] a µ γ TCP By foowing a simiar ine of reasoning, E P can be expressed as foows: = a µ K C, where K C P ac E [N BR] γ TCP. (12) E P = a µ K P, (13) where K P incudes terms that are independent of both a and µ. Deriving the cosed formua of K P requires some manipuation. It can be expressed as K 1 P s + K 3 (P ac P s ) where K 3 = K 2 + K 1 b/bi, K 2 = (E [s] + E [a]) E [N BR ]/MSS, and K 1 = E [N BR ] /γ T CP (see [2, 36]). The inear increase of E C and E P with the average burst size has aso an intuitive expanation. E C is proportiona to the the average downoad interva (E [T], see Equation 11). Assuming E [UTT] = 0, the average downoad interva coincides with the average time spent downoading the bursts from the fixed host, i.e., E [T data ]. Furthermore, since the TCP throughput is assumed to be constant, E [T data ] is proportiona to the average burst size (see Equation 5). In addition, E P is a ineary increasing function of i) the average downoad interva (E [T]), and ii) the average time during which the tagged mobie host remains in the active mode (i.e., E [T ac ], see Equation 11). Based on the above remarks, E [T] is proportiona to the average burst size. Now, we show that the same property hods for E [T ac ], as we. E [T ac ] incudes two components, i.e., the time spent within a downoad interva to receive (transmit) TCP segments (TCP acks), and to receive Beacon frames from the Access Point (see Equation 10). The average tota time required to receive (transmit) TCP segments (TCP acks) is proportiona to the number of TCP segments (TCP acks) managed during the downoad interva, and, hence, to the burst size (Equation 8 and 10). The 15

16 Energy (J) CAM Energy (J) TCP segment oss probabiity Energy (normaized) CAM off scae (150) Energy (normaized) TCP segment oss probabiity TCP segment oss probabiity (a) Energy pots TCP segment oss probabiity (b) Mutipicative factors Figure 7: performance as function of the TCP segment-oss probabiity (p tcp ). tota time required to receive Beacon frames is proportiona to the number of Beacon Intervas within the downoad interva, thus to the downoad interva, and thus to the burst size. The resuts in Figure 6(a) highight an important property of, which is better emphasized in Figure 6(b). Figure 6(b) shows the R(E P, E C ) index 9 as a function of the a parameter. It ceary shows that R(E P, E C ) is amost independent of the average burst size. This is because E P and E C are both proportiona to the burst size, and their ratio depends on the parameters that define K P and K C. In our experiments, this vaue is around 0.16, resuting in an energy saving of approximatey 84%. Based on Figure 6(b) we can caim that the energy saved by does not significanty depend on the average burst size. Therefore, in the foowing experiments, uness otherwise stated, we assume a = Impact of the Internet throughput In this Section we investigate the impact on the performance of the Internet throughput. The resuts presented in [35] show that the segment-oss probabiity (p tcp ) and the average Round Trip Time (RTT) are the main parameters that impact on the throughput of a TCP connection (γ T CP ). Specificay, γ T CP is a decreasing function of both. Thus, we ran a set of simuation experiments to investigate the behavior with respect to p tcp and RTT, respectivey. According to [35], the ower and upper vaues of p tcp are set to and 0.5. E [UTT] is set to 0, as above, whie the rest of the simuation parameters are as in Tabe 1. Figure 7(a) pots E P (bottom curve) and E C (top curve) as functions of p tcp. As expected, both E P and E C increase with p tcp. It is we known that increasing p tcp tremendousy reduces the TCP throughput. The average duration of the downoad interva (E [T]) increases, and this resuts in an increase of both E C and E P. The additiona downoad time mainy consists of onger ide times between burst segments. When is not active, the additiona time is spent competey in the active mode. When is active, the additiona time is ony party spent in the active mode (due to Beaconing), and mosty spent in the seep mode. Hence, is abe to greaty reduce the negative effect of ow throughput on the energy consumption. To quantify this behavior, et us focus on Figure 7(b) that shows the energy consumed at a given segment oss probabiity p tcp, normaized to the energy consumed at p tcp = In a sense, Figure 7(b) shows, for each p tcp vaue, the energy mutipicative factor with respect to the energy consumption at p tcp = For 9 Reca that this index represents the fraction of energy spent when is active, with respect to the energy spent when is not active. Hence, it shows the energy saved by. 16

17 Energy (J) CAM Energy (normaized) CAM RTT (ms) RTT (ms) (a) Energy pots (b) Mutipicative factors Figure 8: performance as function of the Round Trip Time (RTT). exampe, when p tcp is equa to 0.1, the mutipicative factor for E C and E P is around 7x and 3x, respectivey. The mutipicative factor when is active is aways ower than when it is not. Furthermore, the more p tcp increases, the more the difference between the two curves increases. A simiar resut is aso obtained when anayzing the dependence of the energy consumption on RTT (Figure 8(a,b)). Though the absoute vaues are different from those in Figures 7(a) and 7(b), the quaitative behavior is the same. Hence, we can concude that the energy consumption is negativey affected by ow TCP throughput, either is active or not. However, greaty heps in mitigating this effect Impact of the WLAN contention So far, the anaysis has been carried out under the assumption of a singe mobie host within the Wi-Fi hotspot, i.e., M has been assumed to be equa to 0. Now, we evauate the impact of MAC-eve contention on the mobie-host energy consumption (i.e., M > 0). To this end, we firsty highight the imitations of when a standard TCP architecture is used. Then, we investigate up to what extent an Indirect-TCP architecture [11] can aeviate these probems. The simuation parameters are as shown in Tabe 1, apart from E [UTT] which is set to 0, as above in a standard TCP architecture Figure 9(a) pots E P and E C As expected, both E P and E C increase when the contention in the WLAN increases. This behavior stems from two causes: on one hand, MAC-eve contention reduces the TCP throughput; on the other hand, it increases the MAC deay. The impact of the WLAN contention on the TCP throughput ceary appears from Figures 9(b,c,d). The frame oss probabiity on the WLAN increases with M (Figure 9(b)). This resuts in increased number of timeouts at the TCP sender (Figure 9(c)), and, utimatey, to a severe degradation of the TCP throughput (Figure 9(d)). In addition, it is we known that increasing the MAC-eve contention increases the MAC deay [15]. When the MAC deay increases, the time required for receiving a TCP segment (i.e., E [s] in Equation 10), or sending a TCP ack (i.e., E [a] in Equation 10), increases accordingy. So, the time interva during which the tagged mobie host is active (E [T ac ]), and hence E P, increases with M (see Equations 10 and 11). Ceary, simiar remarks appy to E C as we. Based on these remarks, two factors are responsibe for the increased energy consumption when M increases, i.e., i) the reduced TCP throughput (due to an increase in the frame oss probabiity); and ii) the increased MAC 17

18 Energy (J) CAM active stations (M) (a) Energy expenditure WLAN frame oss probabiity active mobie hosts (M) number of timeouts active mobie hosts (M) TCP throughput (Kbps) active mobie hosts (M) (b) WLAN frame oss probabiity (c) number of TCP timeouts (d) TCP throughput Figure 9: performance in a standard TCP architecture. deay. In the foowing of this section, we decoupe the effects of these two factors, to understand the rea impact of each one. Specificay, we show that using an Indirect-TCP architecture [11] eiminates factor i), and expains the discrepancy between the E P and E C curves in Figure 9(a) in an Indirect TCP architecture In an Indirect-TCP architecture [11], the transport connection between the mobie host and the fixed host is spit in two distinct parts at the boundary between the wireess and the wired networks (i.e., at the Access Point). An agent (the Indirect-TCP Daemon) reays the data between the two parts of the connection granting transparency to the appication eve. It has been proved [12] that this architecture shieds the TCP sender at the fixed host from the osses on the wireess ink, thus increasing the throughput with respect to the egacy TCP architecture. We show that this shieding property can be expoited to eiminate the energy wastage reated to the transport appication I TCP Daemon appication STP STP TCP TCP IP IP IP IP MAC MAC mobie host Access Point fixed host Figure 10: Indirect-TCP architecture 18

19 Energy (J) CAM active stations (M) Ideness index active stations (M) (a) Energy expenditure (b) Ideness index WLAN frame oss probabiity active mobie hosts (M) number of timeouts active mobie hosts (M) TCP throghput (Kbps) active mobie hosts (M) (c) WLAN frame oss probabiity (d) number of TCP timeouts (e) TCP throughput Figure 11: performance in an Indirect-TCP architecture. protoco (i.e., cause i) above). To this end, we run simuations by repacing the standard TCP architecture with the architecture shown in Figure 10. This is simiar to the origina Indirect TCP, except for the transport protoco used over the WLAN. Specificay, we use the Simpified Transport Protoco (STP), which is essentiay a Stop-and-Wait transport protoco, optimized for the one-hop wireess environment [6, 3]. Figures 11(c,d,e) show that the Indirect-TCP architecture actuay shieds the TCP sender at the fixed host from frame osses in the WLAN (note that the TCP throughput is measured at the fixed host). Specificay, even though the WLAN frame oss probabiity increases just as in the egacy TCP architecture (compare Figures 11(c) and 9(b)), the number of timeouts registered at the TCP sender (Figure 11(d)) and the throughput experienced by the TCP connection over the wired network (Figure 11(e)) are independent of that. Hence, the effect of the reduced TCP throughput on the energy consumption, registered in the previous set of experiments, disappears. Ony the MACdeay increase (cause ii) above) is thus responsibe for the additiona energy consumption. It shoud be noted that is not abe to face this probem, as it appears from Figures 11(a,b). Figure 11(b) shows the Ideness index as a function of M. The Ideness index is defined as the fraction of time (within bursts) during which the tagged mobie host is ide because there are no frames buffered for it at the Access Point. When the WLAN contention is high (M = 50) the transport-eve throughput on the WLAN is ower than the TCP-throughput on the wired part of the connection. Hence, the TCP sender pumps data towards the Access Point at a higher rate than the tagged mobie host coud fetch from the Access Point. So, the tagged mobie host is never ide, and the can never switch the wireess interface to the seep mode. In concusion, for high contention eves, either enabing the or not eads to simiar resuts (Figure 11(a)). Based on these observations, we can concude that the effect of the MAC deay on the energy consumption can be contrasted ony by reducing the MAC deay itsef through MAC-eve modifications 19

20 (e.g., as proposed in In [14]). To summarize, the resuts presented so far show that in our reference scenario works very we during bursts, i.e., it manages interarriva times very effectivey. Specificay, we have shown that: i) the energy saving achieved by is amost independent of the size of bursts that are downoaded, and, for typica vaues of the main Internet parameters, it can be as high as 84%; and ii) is abe to imit the energy consumption when the throughput offered by the TCP connection drops. 6.3 Is effective with any cass of ide times? Since just expoits the seep mode of the wireess interface, one coud argue that it coud be improved by using the off mode instead. However, this woud cost additiona deay and energy consumption upon re-activation. Whie the transition time from the seep to the active mode (t sa ) is in the order of 1 ms, the transition time from the off to the active mode (throughout referred to as t oa ) is quite greater. The work in [1] measured a transition time around 400 ms, whie [43] measured a transition time around 100 ms 10, which is the vaue we use hereafter. As highighted in Section 6.1, we choose not to focus on the impact of different switching times on energy consumption. Intuitivey, the seep mode shoud be more appeaing for short ide times, whie for ong ide times the best choice shoud be switching the wireess interface off. In this section we corroborate this caim by means of the anaytica mode introduced in Section 5. This suggests some directions to improve the standard. Let us focus on an ide time of a given ength (say, t i ), and et us define the behavior of two idea energy managers, just expoiting the seep and the off mode, respectivey. In the idea case, these energy managers know a-priori the ength of the ide time. The energy manager that uses the seep mode keeps the wireess interface seeping up to t sa seconds before the ide-time endpoint. If E S (t i ) denotes the energy spent by this energy manager during t i, the foowing equation hods: E S (t i ) = (t i t sa ) P s + t sa P ac = t i P s + (P ac P s ) t sa. (14) On the other hand, the idea energy manager that uses the off mode ets the wireess interface in the active mode if t i is ess than t oa. Otherwise, it switches it off, and reactivates it t oa seconds before the ide-time endpoint. If E O (t i ) denotes the energy spent in this case, the foowing equation hods: { t i P ac if t i t oa E O (t i ) =. (15) t oa P ac otherwise Figure 12(a) pots Equations 14 ( idea seep curve) and 15 ( idea off curve) as functions of t i. It confirms that for short ide times the best poicy consists in putting the wireess interface in the seep mode, whie for ong ide times the off-based poicy exhibits the best performance. Let ˆt i denote the crossing point between E S (t i ) and E O (t i ). Then, the optima (idea) poicy is a mixed poicy that uses the seep mode for ide times ower than ˆt i, and the off mode for ide times greater than ˆt i. This anaysis aso suggests that mixed poicies using both the seep and the off modes shoud be defined when short and ong ide times coexist, as in the case of our reference appications. Let us now anayze the energy spent by during t i. Since the station is active just to receive Beacons, the average energy spent by during t i (E P (t i )) is: [ E P (t i ) = t i t ] i BI b P s + t i BI b P ac = t i [ P s + (P ac P s ) ] b BI. (16) Equation 16 is potted in Figure 12(a), with abe. This pot confirms that is effective with respect to interarriva times, i.e., for ide times beow 1 s. The additiona energy expenditure achieved by with respect 10 Actuay, 100 ms is the time measured for a compete cyce active-off-active. Since in our anaysis the breakdown between the active-off and off-active times is not important, we assume 100 ms as the off-active transition time. 20