06 7th Internatonal Conference on Intellgent Systems Modellng and Smulaton A Dynamc Load Balancng Algorthm n Heterogeneous etwork Zhxong Dng Xngjun Wang and Wenmng Yang Shenzhen Key Lab of Informaton Securty and Dgtal Content Protecton Technology Dvson of Informaton Scence and Technology Graduate School at Shenzhen Tsnghua Unversty Shenzhen Chna dzx3@mals.tsnghua.edu.cn wangxngjun@tsnghua.edu.cn yangelwm@mals.tsnghua.edu.cn Abstract Heterogeneous network (Hetet) s one of the key archtectures n 3rd Generaton Partnershp Projects (3GPP) n whch lower-power but more flexble Pco base staton (PBS) s appended. The orgnal homogeneous base staton stays the same and s called Macro base staton (MBS) n Hetet. The tradtonal network admsson and handover s performed by the reference sgnal receved power (RSRP) whch s postvely related to transmsson power. Consequently the lower-power PBS wll be unattractve to user equpment (UE) resultng n feeble performance enhancement of Hetet and mbalance load between MBS and PBS. Ths paper nvestgates a dynamc load balancng algorthm to solve ths load-mbalance problem that deterorate 3GPP LTE network system performance. The algorthm s based on the heterogeneous network access functon (HAF) whch s deduced from network farness analyss accordng to load status of dfferent bases. Meanwhle the algorthm s realzed through traffc transfer scheme and s desgned as a suboptmal but practcal soluton. Smulaton results show that the proposed algorthm gans a better performance on throughput access success rate and the nterlayer farness compared wth exstng scheme. Keywords - Hetet; 3GPP LTE; load balancng; heterogeneous network access functon I. ITRODUCTIO Wth the mprovement of nternet and wreless communcaton technology there s obvous demand growth for hgh data rate servces ncludng multmeda servces hgh qualty of servce (QoS) and abundant network capacty []. Sgnfcant growth also exsts n the amount and dversty of exstng and forthcomng connectng devces ncludng moble phone tablet laptop and ntellgent termnal. Meanwhle devce that s centrally dstrbuted n some hot spots lke bazaar square marketplace and large nner offce space results n nhomogenety of network resource dstrbuton. In order to solve ths problem the Hetet archtecture s ntroduced and also desgned to ncrease network coverage n obstacle areas and orgnal network-edge areas. Hetet s adopted n LTE rado access network (RA) archtecture. And due to ths performance of Hetet the next generaton communcaton system wll also accept t as one basc network archtecture []. The transmsson power of PBS (around 30dBm) s much lower than MBS (around 46dBm) and UE wll connect the base staton (BS) wth max reference sgnal receved power (Max-RSRP) accordng to current communcaton standard [3]. Max-RSRP scheme leads to hgh prorty for UEs to Correspondng author: Dr. Xngjun Wang connect wth MBS resultng n mbalance load between MBS and PBS. Besdes the load-mbalance problem s much more severe because PBS s usually deployed to hot spots wth more UEs. In ths context there have been frutful researches on load balancng whch can be manly dvded nto two aspects: one s on load balancng technology whle the other s on the defnton of network load ndex for UEs to make decsons on whch ths paper focuses. Moreover load balancng technology can be dvded nto channel borrowng scheme (CBS) and traffc transfer scheme (TTS): CBS manly refers to the noton that heavly loaded cell (HLC) borrow unused channel from lghtly loaded cell (LLC) or preempt the shared channel [4]; TTS refers to the noton that UEs are transferred from HLC to LLC wth reasonable allocaton of network resources between MBS and PBS. TTS s more frequently used n ths stage and s also the bass of t research. Ths paper propose a Hetet access functon based on multple load factors n BS and accordngly mplement a suboptmal access control algorthm. We frstly desgn a communcaton smulaton model as a foundaton for the analyss of algorthm then analyze the network farness ndces from whch the Hetet access functon s derved and fnally propose a suboptmal but practcal algorthm after consderng actual mplementaton. In the followng smulaton the proposed algorthm s compared wth the same knds of algorthm to obtan the correspondng conclusons. The rest of ths paper s organzed as follows: Secton II ntroduces related work n load balancng n Hetet. Secton III-A presents the analytcal model for load balancng secton III-B descrbes the network farness ndex and deduce the heterogeneous network access functon and secton III-C ntroduces the dynamc load balancng algorthm and correspondng flow chart. Secton VI evaluates the performance of the proposed algorthm va smulaton at system level and compares wth congener algorthms and secton V draws conclusons. II. RELATED WORK The mult-layer and mult-cell heterogeneous network s faced wth many challenges. We should consder load extent of PBS under coverage of MBS when measurng cell load n mult-layer cell. Subsequently the network load metrc needs to be dstngushed accordng to the base staton type n order to acheve load balancng between MBS and PBS and n MBS and PBS. The author n [6] proposed a dynamc cell- 66-0670/6 $3.00 06 IEEE DOI 0.09/ISMS.06.4 337
sze control (cell breathng) scheme that s an adaptve transmsson power control scheme va whch heavly loaded cell decreases ts coverage and lghtly loaded cells expands ts coverage by adjustng transmsson power. However adjustng transmsson power easly results n network coverage hole whch s hard to optmze [7]. Thus the noton of cell range expanson (CRE) has been ntroduced [8]. In CRE a based receved power value whch s also called range expanson bas (REB) s added to the RSRP of PBS. For example once the REB s assgned the value of 0dB UEs wll access MBS n condton that the RSRP of MBS s hgher than the RSRP of PBS 0dB. However t s dffcult for the fxed REB value to adapt to dynamc heterogeneous networks [9]. It s also hard to understand why UEs should choose PBS wth a lower RSRP uncondtonally. [0] proposed a cell specfc offset to dynamcally adjust REB. Authors n [] used a global convex optmzaton module to analyze load balancng problem workng on optmal REB. But these methods are centralzed mplementatons to acheve optmal performance n overall network wth a relatvely hgh complexty and dffculty to practcally mplement. Our algorthm s based on the local mplementaton to comprse the computatonal complexty and optmal performance as an nstead. III. PROPOSED METHOD A. System model A two-ter heterogeneous network consstng of hghpower MBS and low-power PBS s consdered as shown n Fg.. Fgure. Illustraton of MBS-PBS Hetet The number of MBS s M the number of PBS s P the total number of overall base statons s total and the number of UEs s U. Assume that UEs are evenly dstrbuted n the cell whch s the basc scenaro n the etwork. The dstance between MBS and MBS mantans relatvely certan and further whle the dstrbuton of PBS s stochastc. It s more lkely to appear that the PBS s set up n the network edge or some hot spots whch also can be seen from Fg.. ote that the PBS close to the MBS s more susceptble to hgh-power MBS resultng n UE under the PBS coverage choosng to access to the MBS and thus causng overload n MBS and underload n PBS. Meanwhle UEs n the network edge tend to access PBS because of the use of CRE technology whch causes overload n PBS. The assumpton of these actual scenes shows the necessty for us to take consderaton for load balancng problem n Hetet. Assume that UEs access the cell wth best RSRP sgnals by default. The Max-RSRP scheme can be descrbed as Cell ( u) = arg max RSRP () Max RSRP Su where S u s the set of base statons UE u can receve RSRP sgnals. As n [9-] the noton REB s put forward that the RSRP of PBS s based up wth a certan REB value and the RSRP of MBS remans unchanged. Thus the RSRP-REB can be defned as follows. MCell ( u) = arg max RSRP () Fxed REB S PCell ( u) = arg max{ RSRP + REB} (3) Fxed REB S In the system model the channel state nformaton s perodcally sent back to ts servng cell and the nterference from the same cell s gnored. So the receved sgnal-tonose-plus-nose rato (SIR) of user u from cell g p SIR = u u g ju pj+ 0 j S j p s the transmt power of BS where gus the channel gan between BS and UE u. Thus the RSRP that UE u receve from BS can be calculated as RSRP = g up. 0 represents the nose power that s usually regarded as addtve whte Gaussan nose. The channel gan g u s calculated by the formulaton gu = hh S Lu. h S and h L u represent the small scale fadng factor that s modeled as a Gaussan random varable wth zero mean and unt varance and large scale fadng factor that s denoted as h 0 PL/0 L = where PL s the path loss accordng []. 3GPP LTE system s based on the data servce whch s manly evaluated by the overall throughput. So the Shannon correcton formulaton [3] s used to map the SIR to throughput 0 ( SIRu < SIRmn) Cu = α log ( + SIRu ) ( SIRmn < SIR u < SIR max) C ( SIR max u > SIRmax) (5) where α s attenuaton coeffcent n downlnk channel (We manly analyze the downlnk envronment) SIR s the mnmum target SIR for Adaptve Modulaton and Codng (AMC) max mn (4) SIR s the max SIR when AMC obtans maxmum throughput and throughput. C max s the maxmum 338
B. Hetet Access functon Let s re-consder parameters when UEs access to the BS. Instead of the only parameter of RSRP the actual load condton ncludng the number of UEs throughput n the servng cell and RSRP together wth the farness should be taken nto consderaton. Frst of all there s need to apply a sutable farness evaluaton system n Hetet to assess the overall level of network equlbrum. In terms of farness ndex Jan s farness ndex [4] s wdely used ( X ) = ξ = = where the value of ξ s n [ ]. The value of ξ closer to the more balanced of the network load. When the value s the network load s the most mbalanced that all of network resources are centralzed on one BS cell. The Jan s farness ndex gnores the dfferences of MBS and PBS. There s no doubt that MBS and PBS dffer n network capacty the total UE number and other network resources. Consequently Hetet s regarded as a two-ter network and tered farness ndex [4] s used here β P s used to measure the average layer- network load relatve to the average level of the layer- network load. The β P s defned as follows. M X M M = βp = (7) P X P P = Then the farness ndex s revsed wth tered ndex. X M P ( XM + βpxp ) = = 0 M P total M + βp P = = ξ = ( X ( X ) ) where the value of ξ s also n [ ] total (6) (8) and the expresson can reflect the farness among two layer. To acheve load balancng among two-tered Hetet we need to fnd a sutable ndex when descrbng the degree of resource utlzaton. After consderng several major factors affectng network load ths paper defnes the load ndex as = (9) X ( wu wz) X = X = w U + w Z (0) where X X s U s current served UE rato of BS Z s real-tme traffc represents network load vector and the X absolute value of to the max served UEs number and w w s (throughput) rato to the max throughput of BS. the weght of U and Z and w + w =. w w are assgned as the same value (0.5) n ths paper. The reason why the load ndex s defned as a vector s that dfferent factors nfluencng the network load can be accurately consdered. There s also flexble scalablty that other factors wll be added nto ths load ndex vector lke physcal resource blocks utlzaton and multmeda servce type. Wth the summary of above work the Hetet access functon (HAF) s defned as sgn( X Xcur)sm( X Xcur) G = μ + μ RSRP ξcur () X Xcur sgn( X Xcur) = () ( X Xcur ) ( X Xcur) sm( X Xcur) = (3) X * Xcur where X represents load condton of BS X cur represents average load condton of overall network sgn( X Xcur) s used to get the sgnal of the dfference between X and X cur and sm( X Xcur) s used to characterze the smlarty of X and X cur. ote that the sm( X Xcur) s larger the load gap between X and X cur s smaller. To dscrmnate the load-mbalanced BS we use the square of ξ n the denomnator. And RSRP value s added n the functon snce RSRP s always an mportant condton for UE to access the network normally. The HAF G s used to decde whch cell to access. If G s relatvely large the BS load s lght or the RSRP s large and UE tends to access or vce versa. UE calculates G when choose to access the BS and the global optmum of ξ s converted to dstrbuted local optmum of G wth the computatonal. These centralzed schemes n [0] [] are acheved by brute force search (BFS) wth the computatonal complexty ncreasng as O ( total U ). Ths dstrbuted local sub-optmal scheme can be acheved wth the complexty ncreasng as O (U total ) whch s easer to mplement. 339
C. Dynamc Load Balancng Algorthm Wth summary of the above dscusson a dynamc load balancng algorthm (DLBA) s proposed as shown n Fg. to balance the overall network load n Hetet. and small G. There s no doubt that UEs wll access the BS wth large RSRP and large G and gnore the BS wth small RSRP and small G. But the BSs wth small RSRP but large G and the BS wth large RSRP but small G are hard to dstngush n other conventonal algorthm whch s also a key problem n dynamc load balancng. As for the proposed algorthm DLBA f the BS RSRP s large whle wth a relatvely small G the UE wll gnore t and f the BS RSRP s small whle wth a relatvely large G the UE wll stll access. The network load s balanced through these proposed scheme. And the network throughput s promoted because of the ncrease of UE receved SIR. IV. SIMULATIO A practcal scenaro s smulated to evaluate the performance of our proposed algorthm. Models and parameters are set wth reference to 3GPP standard [3] as shown n Table I. TABLE I. SYSTEM PARAMETERS FOR SIMULATIOS Fgure. The flow chart of DLBA In the proceedngs of the proposed algorthm UEs frstly detect RSRP and judge the type of BS n every cell. If there s only MBS or PBS the dfference between two BS types need not to be consdered. So UE calculates G of neghbor MBS and PBS and correspondngly access the BS wth max G. On the contrary f there are two types of BSs UE preferentally choose the MBS wth max RSRP and the RSRP of ths MBS s supposed as RSRP. UEs sort the servng PBS under the coverage of MBS n descendng order and neglect these PBS j that meets the condton RSRP > RSRPj + REB on whch the UE wll access MBS. For those PBS j meetng the condton RSRP RSRPj + REB UEs calculate the Hetet access functon G and connect the BS wth the max G. Let s relatvely analyze four man state of BS accordng to RSRP and G : small RSRP and small G small RSRP and large G large RSRP and small G and large RSRP Parameter Value Center frequency GHz Bandwdth 0MHz umber of cells 9 macro-cells 3 sectors per macro-cell; 4 low power node-cells per cell Inter ste (MBS dstance) 500m (3GPP Case ) Macro cell coverage radus km Pco cell coverage radus 50m Pco cell placement Random as n 3GPP TS 36.84 [] MBS Tx power 46dBm PBS Tx power 30dBm Shadowng standard devaton Macro: 8dB; Pco: 0dB; As n 3GPP TS 36.84 []: Path loss model 8.+ 37.6log0 d (MBS d n km) 40.7 + 37.6log0 d (PBS d n km) Mnmum SIR threshold -0dB Maxmum SIR for AMC db Measurement nterval 40ms umber of UEs 40 per Macro cell; 0 per Pco cell Alpha (downlnk) 0.6 Maxmum throughput 4.4bps/Hz when SIR s larger than db Channel model TU (ncludng fast fadng) Thermal nose -74dBm/Hz REB value 0 3 6 9 5dB Four PBS are evenly placed n one MBS and we regard one MBS and four PBS as a unt of the regon for measurement (RFM). Smulaton results nclude network throughput network farness and UE access rato. UEs are placed evenly and stochastcally n the Macro and Pco cell complyng wth the number n Table I. It s found that UE number n one Macro cell s 30 accordng to basc scenaro. Wth the regard that the load balancng s a dynamc 340
behavor measurng results durng the dynamc behavor s not accurate and t cannot be compared wth the results of comparatve scheme. As a result a redundancy tme (lke 0 mnutes) s set to guarantee that the system s stable whch s also a proper strategy to avod the cold-start problem. The system-level smulaton s measured by statc snapshot whch s easy to mplement. Besdes ten ndependent smulatons wth random dstrbuton of UEs are performed to work out the average smulaton results. Hetet and therefore can be a proper ndex to descrbe the load n Hetet. ξ of Max-RSRP remans unchanged and ξ of RSRP-REB s equal to the former one when REB s 0 whch valdate the desgn that these two scheme s same when REB s 0. ξ of RSRP-REB and ξ of DLBA get maxmum value around REB = 9dB whle ξ of DLBA s larger and more stable than ξ of RSRP-REB. In order to facltate the followng analyss of other factors we set REB value to 9 db n the next experment. Fgure 3. Rato of accessng MBS under dfferent REB value The rato of accessng MBS n three scheme s shown n Fg. 3. As the value of REB ncreases the MBS accessng rato of Max-RSRP scheme remans unchanged the one of RSRP-REB decreases monotonously and the result of DLBA remans relatvely stable. It can be concluded that Max-RSRP s not sutable to the Hetet and smply addng REB value wll result the mbalance load when REB value s large wth the waste of MBS resources because of the low MBS accessng rato. And the RSRP-REB s hard to dynamcally adapt to the sophstcated and changeable network envronment. The DLBA s an adaptve algorthm and the REB value s not hghly decsve factors n the scheme. The result that rato of accessng MBS and rato of accessng PBS stays relatvely stable valdates the load balancng performance of our proposed algorthm. Fgure 5. Farness ndex under dfferent UE number Fg. 5 ndcates the relatonshp between the farness ndex and UE number that s one of most mportant factor nfluencng load. It s obvous that DLBA has a larger farness ndex than Max-RSRP scheme and RSRP-REB scheme (REB=9dB) of Max-RSRP s relatvely small and decreases slghtly when UE number s too large for UE to obtan suffcent resource of BS. The farness ndex gap between RSRP-REB and DLBA s enlarged and change of RSRP-REB s more sgnfcant. Ths s because MBS reaches the capacty lmt PBS s of relatvely lght load at the same tme and the PBS can t transfer traffc from MBS because of scheme as the UE number ncreases. Whle the same case wll rarely occur n DLBA scheme. Fgure 4. Farness ndex under dfferent REB value We compare Max-RSRP scheme RSRP-REB scheme and our proposed scheme DLBA on the factor of farness ndex as shown n Fg. 4. And revsed farness ndex s used here accordng to Equaton (8) whch s based on two-tered Fgure 6. Throughput under dfferent REB value Fg. 6 shows system throughput n three scheme and local throughput of MBS and PBS n RSRP-REB scheme and DLBA scheme. The reason why the MBS and PBS 34
throughput data of Max-RSRP scheme sn t plotted n the fgure s that most throughput s centralzed n MBS whch s always relatvely unchanged as the REB changes. Throughput of MBS decreases monotonously and the one of PBS ncreases wth REB because UEs wll more nclned to access the PBS under the same condton as REB value ncreases. After syntheszng throughput of MBS and throughput of PBS we get the statstcal data of RFM unt. It s found that ts throughput ncreases as the REB ncreases at frst and then decreases correspondngly. Reasons to explan ths result can be dvded nto two parts: UEs wll start accessng PBS as the ncrease of REB at frst leadng to an mprovement of total throughput; however as the REB value s relatvely large UEs stll access the heavly-loaded PBS whch owns a relatvely small capacty whle the lghtlyloaded but large-capacty PBS s gnored and thus the throughput decreases n ths scene. What s more the throughput gap between DLBA and Max-RSRP or RSRP- REB s enlarged whch s due to the reason that MBS and PBS n DLBA scheme are more fully and effcently used n overall heavy-load stuaton. V. COCLUSIO In ths paper a load balancng problem n 3GPP LTE heterogeneous network s nvestgated n detal. Ths paper frstly researches on how to defne a proper and reasonable scheme for UEs to farly access to Hetet n all ters so as to enhance the user access farness of the overall network and mprove network throughput. A Hetet access functon wth comprehensve consderaton for current channel qualty and the exstng load of neghbor base statons s then proposed. And ths paper fnally makes a tradeoff between complexty and realzablty of the optmum load balancng problem and thus propose a sub-optmal and dynamc load balancng algorthm. Smulaton results valdate the proposed algorthm that t can mprove the balanced load condton farness and system throughput among two ters of Hetet compared wth the Max-RSRP scheme and RSRP-REB scheme. REFERECES [] Csco VI Global moble data traffc forecast update 03-08 Csco whte paper Feb. 04. [] J. Andrews S. Buzz W. Cho S. Hanly A. Lozano A. Soong and J. Zhang What wll 5G be? IEEE J. Select. Areas Commun. vol. 3 no. 6 pp. 065 08 Jun. 04. [3] 3GPP TR 36.84 Further advancements for E-UTRA physcal layer aspects V9.0.0 Mar. 00. [4] SK. Das SK. Sen and R. Jayaram. "A novel load balancng scheme for the tele-traffc hot spot problem n cellular networks" Wreless etworks vol. 4 no. 4 pp. 35-340 998. [5] Alu Osanoh Glenn et al. "A survey of self organsaton n future cellular networks." Communcatons Surveys & Tutorals IEEE 5. (03): 336-36. [6] AM. Sang XD. Wang M. Madhan and RD. Gtln Coordnated load balancng handoff/cell-ste selecton and schedulng n multcell packet data systems Wreless etworks vol. 4 no. pp. 03-0 Feb. 008. [7] Y. Bejerano and S. J. Han Cell breathng technques for load balancng n wreless LAs IEEE Trans. Moble Comput. vol. 8 no. 6 pp. 735 749 June 009. [8] D. Lopez-Perez X. Chu and I. Guvenc On the expanded regon of pcocells n heterogeneous networks IEEE J. Select. Areas Sgnal Process. vol. 6 no. 3 pp. 8 94 June 0. [9] K. Son S. Chong and G. Vecana Dynamcassocaton for load balancng and nterference avodance n mult-cell networks IEEE Trans. on Wreless Communcatons vol. 8 no. 7 pp. 3566-3576 Jul. 009. [0] Lobnger Andreas et al. "Coordnatng handover parameter optmzaton and load balancng n LTE self-optmzng networks." Vehcular Technology Conference (VTC Sprng) 0 IEEE 73rd. IEEE 0. [] Qaoyang Ye and Beyu Rong et al. (03). User Assocaton for Load Balancng n Heterogeneous Cellular etworks IEEE Transactons on Wreless Communcatons 03 pp.706-76. [] 3GPP TS 36.84 Evolved unversal terrestral rado access (EUTRA); Further advancements for E-UTRA physcal layer aspects V9.0.0 Mar. 0. [3] 3GPP E-UTRA; rado frequency (RF) system scenaros 3GPP Tech- ncal Report (3GPP TR 36.94) Dec. 009. [4] DM. Chu and R. Jan "Analyss of the ncrease and decrease algorthms for congeston avodance n computer networks" Computer etworks and ISD systems vol. 7 no. pp. -4 989. [5] MC. Erturk H. Ak I. Guvenc and H. Arslan "Far and QoS-orented spectrum splttng n macrocell-femtocell networks." IEEE Global Telecommuncatons Conference (GLOBECOM 00) Mam Dec. 00. 34