Dmmng Cellular Networks Davd Tpper, Abdelmounaam Rezgu, Prashant Krshnamurthy, and Peera Pacharntanakul Graduate Telecommuncatons and Networkng Program, Unversty of Pttsburgh, Pttsburgh, PA 1526, Unted States Emal: {dtpper, arezgu, prashant, peerap}@ss.ptt.edu Abstract We propose a novel technque called dmmng to mprove the energy effcency of cellular networks by reducng the capacty, servces, and energy consumpton of cells wthout turnng off the cells. We defne three basc methods to dm the network: coverage, frequency, and servce dmmng. We construct a mult-tme perod optmzaton problem to mplement frequency dmmng and extend t to mplement both frequency and servce dmmng together. We llustrate the ablty of dmmng technques to adapt the capacty and network servces n proporton to the dynamc spatal and temporal load resultng n sgnfcant energy savngs through numercal results for a sample network. I. INTRODUCTION Increasng worldwde energy demand, global warmng concerns, and volatlty n energy supples and prces have necesstated mprovng the energy effcency of ICT systems. Recent studes [1], [2] have shown that the largest elements of power consumpton wthn the communcatons component of ICT, are access networks (wred and wreless). The fastest growng access network sector s cellular telephone networks. The ITU estmates that the number of cell phone subscrbers ht 4.6 bllon at the end of 29 [3]. The need for energy effcency n cellular networks has only been recognzed very recently by ndustry, standards groups, and the research communty [1], [2], [4] [1]. Several studes have shown that less than 1% of the energy consumpton n cellular networks s due to moble user handsets whle more than 99% of the energy consumpton s due to the network tself [6]. A study by NTT DoCoMo, Japan s largest moble telecom operator, found that the energy consumpton rato of termnals vs. network s about 1:15 [5]. Wthn the cellular network, base statons (BSs) are responsble for most of the energy consumpton, consumng 5-9% of the network power dependng on the network topology, BS confguraton, rado technology and data rate used [1], [6], [7]. The growth n the number of BSs n the USA llustrates the scale of the ncreasng energy consumpton n cellular networks. The number of cell stes n the USA ncreased from 183,689 n Dec., 25 to 247,81 n Dec., 29 [4]. Wth 3.47 rado base statons on average per cell ste (due to co-locaton and sectorzaton), the number of rado base statons n the USA s closer to 75,, each of whch consumes sgnfcant energy resources. Ths research work s supported n part by a seed grant from the Unversty of Pttsburgh Center for Energy. The work of Peera Pacharntanakul was supported by a predoctoral fellowshp from Offce of the Provost, Unversty of Pttsburgh. Current cellular networks essentally operate n a 24 7 always-on mode consumng sgnfcant power even when not carryng traffc. The network traffc demand however has sgnfcant varatons wth tme of year, tme of day, geographc locaton, weather, socal events, etc. In [1], the dea of powerng down (sleepng) base statons wth no sgnfcant traffc was proposed. However, for reasons such as meetng regulatory requrements, mantanng spatal coverage, synchronzaton, securty and other concerns (descrbed n Secton III), sleepng may not be the rght opton. Thus, the challenge s how to reduce overall energy consumpton n cellular networks whle mantanng adequate coverage, qualty of servces, and relablty. We propose a novel approach to save energy n cellular networks wth dmmng technques where-n one reduces the capacty/servces and thus the energy consumpton of a cell wthout turnng off the cell entrely. BSs can be dmmed n three basc ways. Frst t s possble to just reduce the power of the BS and thereby the coverage area we term ths coverage dmmng. Second, one can swtch off rado frequences (essentally cards n base statons) whch wll reduce the cell capacty, but saves energy by reducng the amount of transmt power and the energy needed for coolng requrements we call ths technque frequency dmmng. Thrd, one can apply servce dmmng where hgh data rate servces are dsabled on a frequency. For a lower data rate, the transmt power s lower and coolng requrements are smaller thus savng energy. In ths paper, we develop network management technques employng frequency and frequency & servce dmmng to save energy n cellular access networks. We consder the spatal and temporal varaton n the network traffc load to determne when and where to employ the dmmng technques whch then dynamcally sze the network and provde only the capacty needed for the gven load n a cell ste. Gven a set of traffc demands, we formulate an optmzaton problem to determne the amount of frequency and servce dmmng employed at each cell. Illustratve numercal results show that ths can result n sgnfcant energy savngs whle stll meetng the traffc demand. The rest of the paper n organzed as follows. In Secton 2, we descrbe a typcal cellular network archtecture and related lterature on energy effcent management technques. Secton 3 contrasts sleepng wth dmmng. Secton 4 presents our optmzaton based formulaton of the proposed dmmng technques. Secton 5 gves numercal results showng the amount of energy savngs possble and
Fg. 2. Representatve SGSN/MSC Load Fg. 1. A typcal cellular network archtecture Secton 6 presents our conclusons. II. BACKGROUND A. Cellular Network Archtecture and Traffc Characterstcs We frst present an overvew dscusson on the current cellular network archtecture n the USA. The Federal Communcaton Commsson regulates spectrum n the USA and has parttoned the country nto 36 Metropoltan Servce Areas (MSAs - whch contan a cty of at least 5, populaton) and 428 Rural Servce Areas (RSAs) for geographc assgnment of the lcenses (map avalable at [11]). A generc 3G cellular moble network archtecture s llustrated n Fgure 1 smlar to a UMTS 3GPP Release 99 network. In UMTS the base staton s termed a Node B and corresponds to a base staton transcever (BTS) n GSM. Smlarly the UMTS Rado Network Controller (RNC) corresponds to a base staton controller (BSC) n GSM and manages a group of base statons and performs rado level channel management and call handoff. The wreless lnk between a Node B and moble s packet based usng wdeband CDMA (WCDMA) as the ar nterface standard. Each Servng GPRS support node (SGSN) detects and regsters new moble termnals n ts servce area, sends/receves data packets to/from the mobles, and tracks the locaton of the mobles wthn ts servce area. The gateway GPRS support node (GGSN) s the nterface between the 3GPP network and external networks. The network may nclude moble swtchng centers (MSC) for connectng to crcut swtched networks. The MSC s connected both to transmsson networks and to the sgnalng network. Assocated wth the sgnalng network and MSC and SGSN are databases to support user and servce moblty (such as the Home Locaton Regster (HLR) and Vstor Locaton Regster (VLR)). Note that the number of cells and ther locaton s normally determned to provde complete geographc coverage whle satsfyng the peak busy perod traffc at that locaton. Gven the cellular network archtecture shown n Fgure 1 consder the network traffc load. Whle cellular voce and data traffc studes have not been wdely publshed due to prvacy and operator concerns, a few have appeared and they clearly show temporal and spatal varatons n the traffc [12] [14]. Fgure 2 shows representatve aggregate data traffc at a SGSN/MSC for an MSA. Note that the traffc has a pronounced durnal behavor whch changes wth day of the week and day of the year, wth the traffc typcally beng much heaver on weekdays. At the level of cells, there are sgnfcant spatal varatons of the magntude and tme when the peak busy perod occurs (ths has the potental for power savngs durng the day as well as at nght). Whle the aggregate traffc would follow the behavor shown n Fgure 2, the traffc at specfc cells can vary greatly. Fgure 3 shows a representatve week of behavor for two cells, one n a downtown area (Cell 1) and the other n a shoppng dstrct (Cell 2). Notce that the traffc load n Cell 1 largely follows the work day. In contrast the traffc at Cell 2 whch provdes coverage to a shoppng/entertanment dstrct has a busy perod much later n the day and skewed more towards the weekend as shown n Fgure 3. One can see that the busy perods are non-concdent. We explot ths later n dmmng cellular networks. Traffc Load (% of Max Capacty) Traffc Load (% of Max Capacty) 8 6 4 2 Cell 1 Load vs. Tme 12 12 12 12 12 12 12 Mon Tue Wed Thu Fr Sat Sun Tme and Day of the week Cell 2 Load vs. Tme 8 6 4 2 Traffc loads n dfferent cells are not algned 12 12 12 12 12 12 12 Mon Tue Wed Thu Fr Sat Sun Tme and Day of the week Fg. 3. Typcal Load at Cells 1 and 2 for Varous Tmes of Day
B. Related Lterature Relatvely few papers have appeared n the research lterature on energy effcent cellular networks [1], [6] [1] Many of the papers are essentally brngng awareness of the need for energy effcency and the challenges n achevng t n cellular networks [1], [5]. Recently, the dea of powerng down base statons wth lttle or no traffc at nght has been advocated [1]. Specfcally, the authors look at the aggregate load on a cellular servce area and based on a rather dealzed analyss show that by turnng off cells to track the tme of day varatons n the aggregate load operators can realze large savngs (25-3%) n energy. However, they do not consder the spatal dstrbuton of traffc n turnng off cells or the addtonal power needed to have neghborng cell expand ther coverage to prevent dead spots n network coverage. A smlar analyss of the wred backhaul porton of the network s gven n [9] showng some possble energy savngs by puttng network components to sleep durng low load perods. Whle the current lterature shows the need for ncorporatng energy effcency n cellular networks and provdes some frst cut analyss of the potental gans, t s lackng n several ways as dscussed next. III. SLEEPING VS DIMMING A. Drawbacks of Sleepng Shuttng down BSs s not straghtforward. Frst, n realty, the traffc has both spatal and tme of day varatons and other factors (weather, holdays, sportng and socal events) affect the traffc load as well. Thus determnng whch cells to swtch off, when and for how long s nontrval and can adversely affect network performance. More mportantly, t s dffcult to envson network operators powerng down cell stes or backhaul equpment entrely for a varety of reasons as dscussed n the followng. Turnng off an entre cell wll result n holes n network coverage, whch can result n poor network performance (e.g., hgh handoff call droppng rate, hgh pagng delays, lack of nterworkng across SGSN boundares for Inter SGSN handoffs, etc.) even though the network load s low. Several papers have appeared n the lterature lookng at the effects of base staton falures on cellular networks (e.g, [15], [16]). One can thnk of a scheduled base staton power down equvalent to a falure. Ths prevous work shows that the mpact of loss of coverage s larger than the geographc area left uncovered. For example, the falure of a BSC n a GSM network knockng out a group of four adjacent cells (n a network of ) results n the mean tme to process a locaton update for the entre group of cells to exceed (by a factor of 1) the recommended ITU benchmark value resultng n protocol tmeouts. Further, the magntude and duraton of the mpact of a loss of coverage (falure) depends on a complex set of factors ncludng the locaton of the loss of coverage (e.g., center or edge of locaton update area, borderng another SGSN/MSC servce area, etc.), shape of the uncovered area (e.g., adjacent or dsjont cells), user moblty patterns, and user behavor n attemptng reconnecton. In fact, the studes show that rado-level falure (e.g., loss of a BS) causes a large ncrease n transent congeston n the sgnalng network [15] due to pagng, dropped handoff and locaton update sgnalng. Another sgnfcant hurdle n turnng off an entre cell ste s that the loss of coverage can result n volatons of regulatory requrements such as the USA E-911 localzaton requrements. The reader s referred to [17] for more detals. Irrespectve of whether AGPS (whch employs assstance from BSs n addton to GPS) or TDOA (BS and network-based trangulaton) postonng s used, the ablty to smply power down BSs s mpacted. Addtonal consderatons nclude ssues as some common cellular equpment needs to mantan synchronzaton (e.g., cdma2 1X-evdo base staton equpment) whch can be slow to reestablsh after beng powered down. Lastly, operators state the need to mantan securty n ther network necesstates keepng t powered up as they are unsure what s happenng to ther equpment when t s powered down entrely. Note that many of these comments apply to the backhaul as well as the wreless part of the network. B. Base Staton Dmmng As dscussed above, the necessty of mantanng geographc coverage, backhaul nterworkng (e.g., Inter SGSN handoffs) and other consderatons (e.g., E-911, synchronzaton, securty) may sgnfcantly constran the number of BSs that can be powered down. In lght of these concerns, we propose a novel approach to energy effcent network management of cellular networks based on the use of dmmng technques where-n one reduces the capacty, servces and energy consumpton of a cell wthout turnng off the cell entrely. Base statons can be dmmed n three basc ways, namely: (1) coverage dmmng, (2) frequency dmmng, and (3) servce dmmng. Coverage dmmng corresponds to reducng the geographc coverage area of a cell, whch s acheved by reducng the base staton operatng power level on all deployed frequences. For example, one can decde to not support soft handoff durng low traffc load perods and reduce the overlappng coverage of cells. Note that the amount of power saved wll be a complex functon of the cell sze, propagaton envronment, lnk power budget and cell confguraton (antennas, locaton, etc). As such t may be dffcult to plan exactly how much to reduce the power and stll have power savngs whle provdng adequate servce and we do not consder ths approach n detal n ths paper. Secondly one can reduce the cell capacty and power use by swtchng off rado frequences, we call ths technque frequency dmmng. Ths saves on both transmsson power and heatng/coolng requrements. Studes have ndcated that up to 4% of the power consumpton of a base staton s n the heatng and coolng. Thrdly, 3G (UMTS, cdma2 1x-EVDO), 3.5G (HSPA) and 4G (LTE, WMAX) networks provde a range of advanced moble data servces (e.g., MMS, streamng vdeo, Internet access, vdeo calls, mult-person gamng etc.) n addton to the 2G servce offerngs of SMS and voce. It s well known that the hgher throughput of the advanced data servces requres hgh power levels (.e., larger E b /N ) f the same coverage footprnt s mantaned. Thus
a power conservng technque for 3G, 3.5G and 4G cells s to consder servce dmmng of operatonal base statons to provde reduced servces (e.g., 2G only). For example, at nght along an Interstate hghway a servce provder can dm cells by dsablng moble data servces and supportng only voce and SMS servces. Note that base staton sleepng s a specal case of coverage dmmng (.e., reduce power to zero) and frequency dmmng (.e., turn off all frequency channels). IV. OPTIMAL DIMMING MODELS Here we consder both the frequency and servce dmmng technques and develop a mult-tme perod optmzaton model formulaton to determne the set of frequences and servces to be offered n each cell as a functon of the load. We frst formulate the frequency dmmng only case and then the combned frequency and servce dmmng scenaro. A. Frequency Dmmng Consder a network composed of a set of N cells provdng servce n a specfc geographc area (.e., MSA or RSA). Let F be the number of frequences lcensed for the servce provder n the geographc servce area. We defne F as the number of frequences assgned to cell. LetP f denote the power consumpton of frequency f n cell where f {1, 2,,F }. Note that P f can be dfferent for each frequency deployed n a cell and ncludes the heatng and coolng power requrements as well as the transmsson power needed. Let C f denote the capacty of frequency channel f at cell. Wedvdeaday nto a set H j of non-overlappng tme perods {1, 2,,H j } on day j, {j =1, 2,, 365}. Note, that the tme perods h {1, 2,,H j } need not be equally long. For the sake of smplcty n the explanaton, we formulate the problem for a sngle arbtrary day and drop the subscrpt for the day of the year. Let W h be the percentage of the day of tme nterval h [.e., (duraton of tme nterval h)/24]. Let d h be the traffc demand at cell durng tme nterval h {1, 2,,H}. We defne x fh as a decson varable where x fh =1f frequency f s used durng tme nterval h at cell and s equal to otherwse. Gven the defntons above, we seek to mnmze the total power PT used n the network whle satsfyng the demand for each tme perod h n a day. Ths can be formulated as optmzaton problem P1 gven below: Problem P1: H N F mn PT = W h P f x fh xfh (1) h=1 =1 f=1 s.t. F f=1 C f xfh d h h, (2) x fh {, 1} h,, f (3) The objectve functon (1) n problem P1 above, seeks to fnd the set of frequences needed to result n the mnmum power needed to carry all traffc demand at each cell n each tme perod n a day. Constrant set (2) ensures that the capacty allocated to a cell for a gven tme perod s greater than the demand. Constrant set (3) ensures the zero or one nature of the decson varables. Note, that to ensure that at least one frequency s left on n each cell thereby mantanng coverage, requres that the d h be set greater than zero for all cells n all tme perods. B. Servce and Frequency Dmmng Gven the optmzaton problem (P1) above we wsh to model the exstence of servce classes and consder the effects of reducng the servces offered on an actve frequency and thereby savng power. Let s {1, 2, S} be the set of servces offered n the network. Servce class s = S represents the hghest class of servce and s =1the lowest. For example an operator mght offer a sute of servce classes on the downlnk such as, HSDPA 3.6 Mbps data (s =4), UMTS 384 Kbps data (s =3), UMTS/GPRS 171 Kbps data (s =2), and UMTS/GSM voce 12.2 Kbps (s =1). Typcally offerng a class of servce mples that all the lower classes of servce are offered as well. For example, f UMTS 384 Kbps class s =3s offered then s =2and s =1are also provded. Note, that f the coverage of the cell s to reman constant then hgher rate data servce requres more transmt power. Let P fs denote the power consumpton of frequency f n cell where f {1, 2,,F } when offerng servce class s where s {1, 2, S}. Then P fs P f(s 1) P f2 P f1. We defne y fsh as a decson varable whch s equal to 1 f frequency f n cell provdes servce class s durng tme nterval h. The basc combned frequency and servce dmmng optmzaton problem can be formulated as optmzaton problem P2 below: s.t. F k=j f=1 Problem P2: Servce and frequency dmmng H N F mn PT = W h P fs y fsh y fsh (4) F C f yfkh f=1 s=1 h=1 C f yfsh F C f yfsh f=1 r=j+1 s=1 d hr =1 f=1 s=1 s=1 d hs h, (5) d hs h, (6) d hj j<s, h, (7) y fsh 1 h,, f (8) y fsh {, 1} h,, f, s (9) The objectve functon (4) n problem P2 seeks to fnd the set of frequences and servces needed to result n the mnmum power requred to carry all traffc demand for each enabled servce class at each cell n each tme perod. The constrant set (5) ensures that the capacty allocated for a cell
Area TABLE I PEAK TRAFFIC AS PERCENTAGE OF CELL CAPACITY Perod (hrs) 7 7 9 9 16 16 18.5 18.5 24 Busness 1 2 9 9 2 Ent. and Shoppng 1 1 35 5 9 Hghway 2 9 2 9 2 Resdental 2 35 2 5 9 TABLE II AVERAGE ENERGY SAVING PER BASE STATION TYPE Cell Type Energy savng Busness 48.33% Ent. and Shoppng 51.67% Hghway 65.% Resdental 44.79% Fg. 4. Sample 5 Node Cellular Network n a partcular tme perod can carry all the traffc demand. Constrant sets (6) and (7) ensure that the enabled capacty for each servce class can carry the traffc demand n that class. Specfcally, constrant (6) requres that the capacty enabled for the hghest servce class n a cell exceed the hghest servce class demand n that tme perod. Smlarly, constrant (7) assures that for the traffc demand n class j, the capacty enabled n the cell for classes j and hgher can carry the traffc for class j and all hgher classes. Constrants (8) and (9) requre that only one servce class s be enabled for a frequency at a tme and the decson varables are requred s 1 f for frequency f servce class s s the hghest servce class enabled durng perod h n cell, and otherwse. Recall that lower servce classes can be supported by a hgher one, e.g., only HSDPA needs to be enabled n order to support UMTS, GPRS, and GSM servces. Optmzaton problems (P1) and (P2) are nteger programmng problems that can be solved usng standard technques such as the branch and bound method. Typcal values of N are 5-2 and we have solved ths problem and smlar ones for ths problem sze usng commercal optmzaton solvers such as CPLEX. However scalablty s an ssue for larger problem szes as nteger programmng problems have NP complexty and heurstcs such as Lagrange relaxaton, genetc algorthms, Tabu search, etc. may be needed for large cases. to be zero or one. Thus, y fsh V. NUMERICAL RESULTS In ths secton, we present numercal results of solvng the base staton dmmng models for a sample UMTS network llustratve of a small cty. The base staton layout conssts of ffty cells (N = 5) as shown n Fgure 4. In the fgure, the cells are classfed nto four categores based on ther traffc profle and the prmary socetal use of the area covered, namely: (1) Busness, (2) Entertanment and Shoppng, (3) Hghway, and (4) Resdental. We dvde a typcal day nto fve tme perods and the peak traffc load n each tme perod n terms of percentage of cell capacty s gven n Table I. We assume the servce provder has a spectrum lcense correspondng to fve FDD 5 MHz UMTS channels and deploys each frequency n every cell. Usng typcal data from UMTS manufacturers product data sheets we set the power consumpton per frequency for HSDPA servce n the busness, entertanment and shoppng, hghway, and resdental areas, as 165, 165, 21, and 2 W, respectvely. Thus a cell n the resdental area usng all fve frequences would consume W of power. Numercal results for the dmmng models are determned usng the branch and bound algorthm mplemented n the optmzaton solver CPLEX 9.1. A. Frequency Dmmng Frst, we present numercal results for the frequency dmmng model (P1). Fgure 5 shows the energy consumpton per tme perod per base staton for each category of base staton (busness, entertanment, hghway, resdental) when frequency dmmng s employed. One can clearly see from the fgures that substantal energy savngs are possble from frequency dmmng. Computng the average daly energy savngs n each traffc area we get the results shown n Table II. Notce that energy savngs up to 65% are found n portons of the network. B. Frequency and Servce Dmmng To evaluate the effects of combned servce and frequency dmmng, we determned numercal results for the same ffty node network. We assumed four classes of servce are provded at each cell (HSDPA data, UMTS data, GRPS data, UMTS/GSM voce). We approxmated the energy savngs per frequency by provdng only lower class servces by gatherng data from UMTS base staton manufacturers product data sheets. Specfcally the power savngs per frequency wth the reducton n servce from HSDPA to UMTS, GPRS, and GSM are 2%, 25%, and 35%, respectvely. The total load per category of base staton follows the same tme of day varatons gven n Table I. However, the mx of traffc among the four servce classes for each tme perod also vares wth tme of day. In Table III, we show the traffc demand per servce class as a percentage of the maxmum capacty of the base staton. The demands are gven as a four-value tuple n decreasng order of servce class (HSDPA, UMTS, GPRS, GSM).
5 5 Traffc profle: Busness dstrct 1 2 3 4 5 6 7 8 9 1 11 12 13 14 15 16 17 18 19 2 21 22 23 24 Traffc profle: Ent&Shoppng dstrct 1 2 3 4 5 6 7 8 9 1 11 12 13 14 15 16 17 18 19 2 21 22 23 24 Traffc profle: Hghway dstrct 5 1 2 3 4 5 6 7 8 9 1 11 12 13 14 15 16 17 18 19 2 21 22 23 24 Traffc profle: Resdental dstrct 5 1 2 3 4 5 6 7 8 9 1 11 12 13 14 15 16 17 18 19 2 21 22 23 24 5 Traffc profle: Busness dstrct 1 2 3 4 5 6 7 8 9 1 11 12 13 14 15 16 17 18 19 2 21 22 23 24 Traffc profle: Ent&Shoppng dstrct 5 1 2 3 4 5 6 7 8 9 1 11 12 13 14 15 16 17 18 19 2 21 22 23 24 Traffc profle: Hghway dstrct 5 1 2 3 4 5 6 7 8 9 1 11 12 13 14 15 16 17 18 19 2 21 22 23 24 Traffc profle: Resdental dstrct 5 1 2 3 4 5 6 7 8 9 1 11 12 13 14 15 16 17 18 19 2 21 22 23 24 Fg. 5. Percentage of Maxmum Power Consumpton Vs. Tme for Varous Types of Cells wth Frequency Dmmng Area TABLE III PEAK TRAFFIC PER SERVICE CLASS FOR EACH TIME PERIOD AS PERCENTAGE OF THE CELL CAPACITY Perod 7 7 9 9 16 16 18.5 18.5 24 Busness,,,1,1,5,5 3,25,2,15 3,25,2,15,,,2 Ent. and Shoppng,,,1,,,1,5,15,15 1,,2,2 3,25,2,15 Hghway,,1,1 3,25,2,15 5,5,5,5 3,25,2,15,,1,1 Resdental 5,5,5,5 5,1,1,1 5,5,5,5 15,1,15,1 3,25,2,15 format: HSDPA, UMTS, GPRS, GSM TABLE IV AVERAGE ENERGY SAVING PER BASE STATION TYPE Cell Type Energy savng Busness 58.65% Ent. and Shoppng 62.42% Hghway 7.6% Resdental 51.33% Fgure 6 shows the energy consumpton per tme perod per base staton for each category of base staton (busness, entertanment, hghway, resdental) when both frequency and servce dmmng optons are utlzed. Table IV lsts the average daly energy savngs for each base staton type. One can see that servce dmmng provdes addtonal energy savngs on top of frequency dmmng. VI. CONCLUSION In ths paper, we propose the novel dea of dmmng cellular networks for energy savngs. We explore frequency dmmng whch swtches off certan frequency carrers n cells and servce & frequency dmmng that dsables certan hgh data rate servces on enabled frequences durng perods where traffc demand s low. We formulate a framework for optmally dmmng cells and show that up to 65% savngs n average energy consumed n parts of a network s possble wth frequency dmmng and up to 7% savngs n average energy consumed n parts of a network s possble wth both frequency and servce dmmng. Fg. 6. Percentage of Maxmum Power Consumpton Vs. Tme for Varous Types of Cells wth Frequency and Servce Dmmng ACKNOWLEDGMENT The authors would lke to thank Mr. Tae-hoon Km for hs help wth some of the llustratons n ths paper. REFERENCES [1] G. Fettwes and E. Zmmermann, ICT energy consumpton trends and challenges, The 11th Internatonal Symposum on Wreless Personal Multmeda Communcatons (WPMC), 28. [2] ITU and clmate change. [Onlne]. Avalable: http://www.tu.nt/themes/clmate/ndex.html [3] ITU, The World n 29: ICT Facts and Fgures. Geneva, Swtzerland: ITU Telecom World, October 29. [4] CTIA: The wreless assocaton. [Onlne]. Avalable: http://www.cta.org [5] Green moble networks and base statons: Strateges, scenaros and forecasts 29-214, Junper Report, Tech. Rep., July 29. [6] M. Etoh, T. Ohya, and Y. Nakayama, Energy consumpton ssues on moble network systems, Internatonal Symposum on Applcatons and the Internet (SAINT), 28. [7] D. Lster, An operator s vew of green rado, keynote speech, n Proc. IEEE GreenCom, June 29. [8] Green communcatons, ICC 29 Panel P4, IEEE Internatonal Communcatons Conference, June 29. [9] L. Charavglo, M. Mella, and F. Ner, Energy-aware umts core network desgn, The 11th Internatonal Symposum on Wreless Personal Multmeda Communcatons (WPMC), 28. [1] M. Marsan, L. Charavglo, D. Cullo, and M. Meo, Optmal energy savngs n cellular access networks, n Proc.IEEE GreenCom, Jun 29. [11] FCC wreless telecommuncatons bureau cellular market areas. [Onlne]. Avalable: http://wreless.fcc.gov/auctons/data/maps/cma.pdf [12] E. Halepovc, C. Wllamson, and M. Ghader, Wreless data traffc: A decade of change, IEEE Network, vol. 23, no. 2, March 29. [13] T. Wnter, U. Turke, and M. Koonert, A generc approach for ncludng lve measurements and traffc forecasts n the generaton of realstc traffc scenaros n moble rado networks, ACM Internatonal Workshop on Modelng Analyss and Smulaton of Wreless and Moble Systems (MSWM 24), Oct 24. [14] C. Wllamson, E. Halepovc, H. Sun, and Y. Wu, Characterzaton of CDMA2 cellular data network traffc, Proc. IEEE LCN, Nov. 25. [15] D. Tpper, C. Charnsrpnyo, H. Shn, and T. Dahlberg, Survvablty analyss for moble cellular networks, CNDS, 22. [16] A. P. Snow, U. Varshney,, and A. D. Malloy, Relablty and survvablty of wreless and moble networks, IEEE Computer, vol. 33, pp. 49 55, 2. [17] Y. Zhao, Standardzaton of moble phone postonng for 3g systems, IEEE Communcatons Magazne, pp. 18 116, 22.