Cyber Journals: Multdscplnary Journals n Scence and Technology Journal of Selected Areas n Telecommuncatons (JSAT) February Edton 2013 Volume 3 Issue 2 Load Balancng Based on Clusterng Methods for LTE Networks Omar Altrad Member IEEE Sam Muhadat Senor Member IEEE Abstract In ths paper we propose a general load-balancng algorthm to help congested cells handle traffc dynamcally. The algorthm s based on clusterng methods and can be appled to any wreless technology such as LTE WMAX and GSM. The algorthm can be automatcally controlled and trggered when needed for any cell on the system. It can be mplemented n a dstrbuted or sem-dstrbuted fashon. The trggerng cycle for ths algorthm s left for the operator to decde on; the underlyng varatons are slow so there s no need for fast self-optmzng network (SON) algorthms. We apply the load-balancng algorthm to an LTE network and dfferent crtera are adopted to evaluate the algorthm's performance. T Index Terms About Load balancng LTE Handover SON. I. INTRODUCTION HIS load experenced by neghborng cells tends to vary dependng on the tme of day and centers of actvty; ths causes cells to be more or less congested. Dfferent dstrbutons of traffc occur n both space and tme whch leads to unbalanced loads n the cells and causes degradaton n system performance. Ths temporary traffc concentraton problem needs a dynamc mechansm to adapt for these changes ether by usng more hardware resources or the careful desgn of an algorthm to treat these occurrences. The load-balancng algorthm ams to fnd the optmum handover offset (HO) value between the overloaded cell and a possble target cell. The use of load-balancng (LB) whch belongs to the group of suggested SON functons for LTE network operatons s meant to delver ths extra gan n terms of network performance. In addton the algorthm needs to adust the network control parameters n such a way that overloaded cells can offload the excess traffc to low-loaded adacent cells [1] In [2] a method of balancng the load among cells whch are operatng at maxmum capacty s descrbed. However ths method has the dsadvantage of handlng the handover of the moble staton (MS) due to load balancng dfferently from the handover of the MS leavng the cell. Another approach [3] patented by Koma narrows the servce area of the base staton (BS) by reducng ts output power. However ths Manuscrpt receved January 30 2013. O. Altrad and S. Muhadat are wth the School of Engneerng Scence Smon Fraser Unversty Burnaby BC V5A1S6 Canada. Phone: 778-782- 7376. Fax: 778-782-4951. E-mal: oaltrad@sfu.ca muhadat@eee.org 1 patent does not dscuss the manner n whch the handng over of establshed calls takes place. Bodn [4] ntroduces the concept of adaptve handover boundares and ntroduces a smple algorthm to solve ths problem. However ths algorthm does not ensure the exstence of a contnuous overlappng area. In [5] a soluton for the adaptve handover problem s consdered based on the predctable pattern of traffc loads; however ths assumpton becomes neffcent when a devaton between the current pattern and the analyzed hstorcal traffc patterns occurs. Most of the prevous work on evaluatng the performance of load-balancng algorthms for cellular networks emphaszes smulatons [6] [7]. Other papers adopt the theoretcal analyss approach whch nvolves usng mathematcal technques such as queung models and Markov chan models to model and study the performance of task schedulng algorthms [8] [9]. Our contrbuton n ths paper s the followng: We ntroduce a new load-balancng algorthm based on clusterng methods the centrod of the cluster s the cell poston; a mathematcal formulaton for the problem to analyze the algorthm s ntroduced; The trggerng mechansm for the algorthm s the call blockng rato (CBR) whch s the real parameter reflectng the degradaton of the system when overload occurs; a control functon s ntroduced and mplemented wth the proposed message names for reducng sgnalng overhead between the cells. The rest of the paper s organzed as follows. A descrpton of the algorthm and ts behavor s dscussed n secton II. In secton III a mathematcal analyss of the algorthm s presented. Secton IV provdes the smulaton results and the paper s concluded n secton V. II. ALGORITHM DESIGN AND DESCRIPTION The proposed algorthm s sem-dstrbuted snce t s nvoked at each congested cell and controlled by a management entty. The nput for ths algorthm s the current load of each neghborng cell as well as the current handover margn whch s shown n (11). The proposed algorthm works as follows: Frst each moble staton reports ts measurements to ts servng cell n a perodc fashon. These measurements nclude the SNIR measurements of the neghborng cells as well as the servng
cell. At any tme and for any cell n the system f (12) s satsfed then the cell s consdered to be congested and a REQ_Load_Balance message s sent to the management entty. Ths wll respond by sendng a Load_Balance_Res message to nvoke the algorthm on the requested cell and at the same tme ths cell s added to the lst of the update cell pool. We need ths process before startng the algorthm n any cell n order to allow the management entty to exclude ths cell after t has fnshed runnng the algorthm so as to prevent an endless loop n the system. Also note that t s straghtforward to modfy the algorthm to be executed wthout the management entty.e. fully dstrbuted. However the drawback of ths wll be extra sgnalng overhead between cells as well as an ncreased png-pong effect snce every congested cell wll try to shft the overload to ts neghborng cells when congeston occurs n more than one neghborng cell. After that the congested cell wll request the estmated load of each neghborng cell n the lst. Ths reflects the current load of the cells. Each cell wth an estmated load less than the defned threshold load wll be added to the canddate lst to be assgned a porton of the load of other cells wth the condton that ths cell s a neghborng cell of the congested cell. Note that we have only one threshold defned to dstngush between congested and decongested neghbors. Ths threshold s enough to do the ob snce the algorthm wll recursvely choose the next congested cell and exclude the cells that are already fully loaded. A mappng functon s used to update the handover threshold of the overloaded cells as shown n (15). The algorthm s smple and the only requrement s the measurement exchange of the estmated load. Ths extra sgnalng depends only on the sze of the lst of neghbors and the perodcty rate. It s not; however approprate to substantally decrease the sze of the lst of neghbors due to handover optmzaton ssues. The followng scrpt s used to llustrate the proposed algorthm. Proposed Algorthm A moble staton reports ts measurements to the servng cell usng (5). The servng cell Req_Load_Balance message. The management entty adds cell by sendng Load_Balance_Res. detects an overload usng (12) and sends the to the update balance pool and responds FOR each cell (congested cell) the clusterng functon s nvoked to estmate the overloaded porton wth the constrant of (10) FOR each cell neghbor to = 1 k performs Update HM (; ) usng (15) so as to reduce the coverage area whch depends on the avalable resources of ts neghbor. = + 1 END Informs the management entty by sendng Update_Balance_ Fn. The management entty drops cell from the balance pool f Update_Balance_Fn s receved. = + 1 END III. ALGORITHM ANALYSIS Consder a cellular coverage area C consstng of n cells C{ C C... C } 1 2 statons defned as Denote M 1... m as the moble staton n 1... n and a set of all moble M M1 M2 M m {... } connected to C and m n. Then the receved power at the moble staton from base staton defned as Where t P PG t r M L P s the transmtted power of the cell the gan between the moble and L l d frequency beng used. G and the cell. s (1). Where l s constant depends on the d s s the dstance between the moble and the cell. s the path loss exponent and represents the shadowng effect whch can be modeled as shown n [10]. The measured sgnal to nterference and nose rato at N X and M can be defned as SINR M I P r M N s the thermal nose and I s defned as X ( p) p p I X p P p r p ( ) M s defned as 1 when p use the same band 0 when p use dfferent band s defned as the load rato of used resources. We represent each M (2) (3) (4) as a pont n space of a dmenson equal to the length of the neghborng base statons lst and the servng cell.e. D f SINR SINR SINR (5) M 1 k {1 k} represents the length of the lst of neghborng cells. Usng ths conventon we can represent the cell C as a pont n space of the same dmenson.e. D C f SINR1 SINRk SINR (6) 2
We assume that the cell can measure the receved power of ts neghbors as t wll be the centrod of the clusters. Then the Eucldan dstance between M and X D D M M C C wll be The clusterng algorthm must be appled recursvely so as to map the load n each congested cell to a number of clusters and then use ths mappng to adust the handover margn wth each neghborng cell. By default ths prevents or delays the headng MSs from connectng to the congested base staton or extendng the connecton of an MS n a lght-loaded cell to a certan lmt. To apply the clusterng method some preparaton s needed and a number of requrements must be met for ths process to be accomplshed successfully. We can represent the overall load of the congested cell as L C 2 S K 1 S 1 represents the frst cluster sze whch s ntended to be handed over to neghborng cells. K S 2 (7) (8) represents the second cluster sze whch represents the acceptable load of can be handled.e. t can be constraned by L th S L 2 th C that (9) s a predefned threshold for each cell whch represents the maxmum allowable load on the cell. Note that ths threshold does not mean we reserve resources n the cell snce t can smply be replaced by the maxmum allowable load n that cell. However we defne t here as such for llustraton purposes. Estmatng the sze of one cluster wll gve us the sze of the other cluster. To do so we sort the moble statons accordng to ther Eucldan dstances from the congested cell and keep addng the requested resources for each moble tll we reach the maxmum load threshold that can be handled by the cell. Ths dstance wll ndcate the crossover pont between the two clusters. The sze of cluster S 1 can be constraned by the sum of all avalable resources n the neghborng cells: Where k S1 (1 L ) (10) 1 L s the estmated load of the neghborng cell. Equaton (10) s mplemented to prevent an endless loop between cells and to reduce the png-pong effect when the algorthm s executed. Note that an entre cluster s not necessarly handed over to one neghborng cell only snce each moble staton n the cluster wll be handed over to a preferred cell ndcated by the best SNIR receved.e. cluster S1 can be consdered to be dvded nto sets; each set wll be connected to the neghborng cell wth the best SNIR receved. For more clarfcaton of ths pont see Fg.1 we consder two neghborng cells for llustraton purposes. A. Handover Trggerng Condton The condton whch trggers the handover from a servng cell to ts neghborng cell can be dependent on many factors. Some examples of these are BER SINR and RSSI. In LTE networks the hard handover algorthm or so called Power Budget Handover Algorthm s adopted. Two parameters are defned n the cell at the tme of deployment to swtch the moble user from one cell to ts neghbor the handover margn (HM) and the tme to trgger (TTT). These parameters are constant and mplemented durng the deployment phase. Dfferent values are consdered for each cell on the system. The receved sgnal strength s called the reference sgnal receved power (RSRP) whch s used to evaluate whether the condton to trgger a handover has been met. Ths condton can be wrtten as. RSRP RSRP HM T RSRP RSRP T S S (11) are the reference sgnal receved power of the targetng cell and the servng cell respectvely. Ths condton must be satsfed for a perod of tme represented by the TTT. B. Detecton of overloaded cells and handover adaptaton Before nvokng the algorthm a trggerng method should be used to detect the overloaded cells. We used CBR as the trggerng method.e. when B th CBR B th CBR blocked calls / total accepted calls (12) and s a predefned threshold kept for operator use whch s determned by the qualty of servce (QoS) the operator promsed to provde. In ths paper we kept ths threshold to 2% [11]. A mappng functon to adust the handover margn between each par of cells should be used. The adusted handover margn between the congested cell and ts neghbor wll be drectly proportonal to the estmated load n the neghborng. The overloaded porton represented by the cluster S 1 defned n (10) can be dvded nto subsets; each subset wll reflect the amount of load ntended to be handed over to a neghborng cell. Ths subset wll be constraned by the maxmum allowable load that can be shfted to ths neghborng cell.e. S { s s s } 1 11 12 1 k (13) s1 s the subset of neghbor cell. The sze of the subset can be constraned by s L (14) 1 1 The handover margn between the congested cell and the 3
neghborng cell wll then be adusted as ( ) HM HM HM H s def def max (15) Note that the adustng procedure requres only one step. Ths method wll dramatcally reduce the sgnalng overhead caused when compared to conventonal adustment methods. The adopton of the lnear equaton shown n (15) s because the handover margn s drectly proportonal to the overload porton. It thus follows that a lnear functon s suffcent for ths procedure. Equaton (15) s a smart way of adustng the handover margn when the cells are congested snce ths adustment s par-wse adusted.e. each neghbor has ts own adusted handover margn wth the congested cell. The algorthm we propose wll handle ths adustment procedure. IV. SIMULATION RESULTS For the performance evaluaton of the proposed algorthm a modfed LTE model based on Opnet modular 16 smulaton software s adopted as shown n Fg.2. A scenaro that reflects the practcal stuatons of MSs movement and envronment s modeled. Table 1 PARAMETERS VALUES FOR SIMULATIONS Attrbute Value Base Frequency 2 GHz Network layout 7 BS ste 3 sectors 21 cells Path loss Model Okumura-Hata COST 231 Bandwdth 10 MHz Bandwdth 10 MHz Maxmum transmsson power 46 dbm Physcal Profle Type OFDM Moblty model Random Waypont Model Thermal nose (N) -114 dbm Shadowng zero mean and standard devaton 8 db Symbol Duraton 100.8 mcroseconds Number of subcarrers per RB 12 Sub-carrer Spacng 15 KHz Packet Scheduler Round Robn Inter ste dstance 500 m Threshold load Lth 0.85 of the maxmum estmated load Frame Duraton 10 mllseconds A. Scenaro and Parameters The maor smulaton parameters are shown n Table I we follow the reference settngs for LTE [12]. Each BS ste has three sectors; each sector represents a cell of hexagonal shape. A total of 21 cells are used n ths study. The MSs are unformly dstrbuted n the lghtly loaded cells. Each MS s constantly movng at a fxed speed and wth an ntal drecton randomly chosen from 0 to 2 the MS s permtted to change ts drecton randomly so as to represent practcal stuatons. Moreover we have created a cluster of MSs havng random movement whch can be dropped at dfferent tmes of the smulaton nto randomly chosen cells to represent buses and trans. The Costa-231 HATA model for the urban envronment s used for the path loss computaton [13]. Shadow fadng s modeled as a Gaussan log normal dstrbuton wth 0 mean and 8 db varance [14]. A Round-Robn packet scheduler s used for far transmsson whle the Hybrd Automatc Repeat Request (HARQ) technque s used for wreless transmsson error recovery [15]. Assumng a constant bt rate of 256 kbps for each user and a bandwdth of 10 MHz ths wll cause an approxmately 38 users/cell. The algorthm descrbed earler wll adust the handover threshold for each congested cell wth ts neghborng cells whch requres only one step as dscussed earler. Startng from the most congested cell the adustng procedure wll follow the Eucldan dstance mappng descrbed earler. 10 db s the maxmum handover adustable margn. The default value s kept unchanged when there s no congeston n the cell the maxmum means the cell s fully loaded. All cells of stes BS_2 BS_4 are consdered as congested cells wth an average 42 users/cell. All other cells belongng to stes BS_1 BS_3 BS_5 BS_6 and BS_8 are lghtly-loaded cells wth an average of 12 users/cell. B. Results In the smulaton scenaro we compare the performance of the proposed system when mplementng the load-balancng algorthm and when not nvokng the algorthm. The performance s measured n terms of CBR. Snce ths algorthm entals only one-step adaptaton the result was promsng and the CBR was dropped to almost zero n all congested cells as shown n Fg. 3. For example before the algorthm was nvoked the CBR n cells 7 8 and 9 was 12% 9% and 10% respectvely whch contrbuted to a CBR average of 10% n the three congested cells. Ths congeston was reduced to almost 1% when the algorthm was appled. Thus a reducton n congeston of 90% was acheved. Also notce that the average load of cell 14 exceeded 2 ths load represents the blocked and accepted calls as shown n Fg. 4. At the same tme the CBR of ths cell was 10% as shown n Fg. 3. Rght after nvokng the algorthm as seen n Fg. 4 the load on cell 14 was reduced to the threshold 85% whch we defned earler. Ths reducton of load n the congested cell was carred by cell 3 and cell 5 each cell carred a porton of the overload. Note that cell 15 and cell 13 dd not handle any porton snce they were already congested cells. Also notce that even though cell 11 s a lghtly-loaded cell t dd not contrbute to the process as we lmted the lst of neghbors to only three to reduce the smulaton tme. Fg. 5 shows a comparson between the algorthm suggested by [4] and our proposed algorthm. The Bodn algorthm was mplemented after some modfcaton to ft LTE requrements. The algorthm proposed by [4] dd not show the step sze or the way the load of the congested cell was estmated. Moreover t dd not explan the nvokng procedure or the message exchange between cells. As a result our proposed algorthm showed a reducton of more than 80% n CBR compared to the Bodn algorthm. Ths reducton was also caused by the fne adustment of our proposed algorthm and 4
the control procedure we adopted. Moreover n Fg 6 a fluctuaton of the load n the congested cell s seen when applyng the Bodn algorthm whch explans the hgh CBR and the nstablty of the algorthm. Our proposed algorthm on the other hand shows consstency and a smooth control of the current load of the congested cell. Fnally one of the most mportant features for any proposed load-balancng algorthm s fast adaptablty to the dynamc changes n the load n the congested cells whch results n a reduced CBR. Therefore the number of satsfed users s ncreased. Ths feature s shown n Fg. 6 the convergence of our proposed algorthm s acheved wth less tme compared to the Bodn algorthm. V. CONCLUSION In ths paper a load-balancng algorthm based on clusterng methods s proposed. We appled ths algorthm to LTE networks. Our results show a sgnfcant mprovement compared to prevous works. Usng a par-wse method to adust the handover margn sgnfcantly mproves the performance of the system compared to the conventonal methods whch use the cell-wse method. Our new method shows a reducton n the CBR exceedng 85% n some cells. Moreover a total reducton of 75% n CBR s acheved on the overall system. Our results show a dstrbuton of the load of the congested cell to ts neghbor n one step only whch sgnfcantly reduces the sgnalng overhead and wastng of resources n the lghtly-loaded cells compared to conventonal methods. Applyng ths algorthm to more practcal scenaros and relaxng some of the assumptons made here s left for future work. VI. REFERENCES [1] A. Lobnger S. Stefansk T. Jansen and a. I. Balan "Load balancngn downlnk LTE self-optmzng networks" n Vehcular Technology Conference (VTC 2010-Sprng) May 2010. [2] Brody and C. George "Load balancng for cellular radotelephone system". US Patent 4 670 899 2 January 1987. [3] J. Koma and K. Mzoe "Rado moble communcaton system n probablty of loss of calls s reduced wthout a surplus of basestaton equpments". US Patent 4 435 840 6 March 1984. [4] R. bodn Spanga and A. Norefors "Load sharng control for a moble celluar rado system". US Patent 5 241 685 1 August 1993. [5] C. Chandra T. Jeanes and W. Leung "Determnaton of optmal handover boundares n a cellular network based on traffc dstrbuton analyss of moble measurement reports" n Vehcular Technology Conference 1997. [6] T. Nhtla J. Turkka and I. Verng "Performance of LTE selfoptmzng networks uplnk load balancng" n Vehcular Technology Conference (VTC Sprng) May 2011. [7] J. Rodrguez I. d. l. Bandera P. Munoz and R. Barco "Load balancng n a realstc urban scenaro for lte networks" n Vehcular Technology Conference May 2011. [8] I. Verng M. Dottlng and A. Lobnger "A mathematcal perspectve of self-optmzng wreless networks" n Internatonal Conference on Communcatons (ICC) June 2009. [9] S. Kourts and R. Tafazoll "Adaptve handover boundares: a proposed 5 scheme for enhanced system performance" n Vehcular Technology Conference Proceedngs 2000. [10] O. Altrad S. Muhadat and M. Danat "A novel-dual trgger handover algorthm n wmax networks" n Internatonal Wreless Communcatons and Moble Computng Conference(IWCMC) Aprl 2012. [11] A. Technologes "Aglent 3GPP Long Term Evoluton: System Overvew Product Development and Test Challenges." Avalable onlne: http://cp.lterature.aglent.com/ltweb/pdf/5989-8139en.pdf 2009. [12] 3GPP "Physcal Layer Aspects for evolved Unversal Terrestral Ra-do Access (UTRA)" 3GPP. [13] T. S. Rappaport Wreless Communcatons: Prncples and Practce 3rd ed. Prentce Hall 2003. [14] M. Gudmundson "Correlaton model for shadow fadng n moble rado systems" Electroncs Letters vol. 27 no. 23 pp. 2154-2146 1991. [15] J. Ikuno C. Mehlfuhrer and M. Rupp "A novel lnk error predcton model for OFDM systems wth HARQ" n Internatonal Conference on Communcaton (ICC) 2011. Unversty Canada. Omar Altrad (M 09) was born n Irbd Jordan n 1974. He receved the B.Sc. n communcaton engneerng from Mutah Unversty Jordan 1996 and the M.sc. degree n electrcal and computer engneerng from New York Insttute of technology USA. He worked as a feld engneer Drector of staff at Royal Jordanan Ar Force (1996-2006). Currently he s pursung hs Ph.D. n wreless communcatons at the school of engneerng Scence Smon Fraser Frst A. Author (M 05 SM 12) receved hs M.Sc. n Electrcal Engneerng from Unversty of Wsconsn Mlwaukee USA n 1999 and the Ph.D. degree n Electrcal Engneerng from Unversty of Waterloo Waterloo Ontaro n 2006. From 1997 to 1999. he worked as a Research and Teachng Assstant n the Sgnal Processng Group at the Unversty of Wsconsn. From 2006 to 2008 he was a postdoctoral fellow n the Department of Electrcal and Computer Engneerng Unversty of Toronto. From 2008 tll present he s assstance professor n Engneerng school Smon Fraser Unversty.
Fg. 4 Tme average load n the congested cell 14 and ts lghtlyloaded Fg. 1 Illustraton of the subset for cluster S1 Fg. 5 Call blockng rato of our proposed algorthm Compared to Bodn Fg. 6 Tme average load of our proposed algorthm Compared to Bodn n Fg. 2 Network Layout Fg. 3 Call blockng rato for each cell n the network 6