1 Dynamic Sounding for Muli-User MIMO in Wireless LANs Xiaofu Ma, Suden Member, IEEE, Qinghai Gao, Ji Wang, Vuk Marojevic, and Jeffrey H. Reed, Fellow, IEEE Absrac Consumer elecronic (CE) devices increasingly rely on wireless local area neworks (WLANs). Nex generaion WLANs will coninue o exploi muliple anenna sysems o saisfy he growing need for WLAN sysem capaciy. Mulipleinpu muliple-oupu (MIMO) anenna sysems improve he specral efficiency and single user hroughpu. Muli-user MIMO (MU-MIMO) sysems exploi he spaial separaion of users for increasing he sum-hroughpu. In an MU-MIMO sysem, efficien channel sounding is essenial for achieving opimal performance. The sysem analysis in his paper provides insighs ino he rae a which o perform channel sounding. This paper shows ha opimal sounding inervals exis for single user ransmi beamforming (SU-TxBF) and MU-MIMO, and proposes a low-complexiy dynamic sounding approach for pracical MU- MIMO WLAN deploymens. The proposed approach adjuss he sounding inerval adapively based on he real-ime learning oucomes in he given radio environmen. Using real over-he-air channel measuremens, significan hroughpu improvemens (up o 31.8%) are demonsraed by adoping he proposed dynamic sounding approach, which is complian wih IEEE 82.11ac 1. Index Terms Muli-user MIMO, wireless local area nework, dynamic sounding, channel sae informaion. I. INTRODUCTION Wireless local area neworks (WLANs) have experienced dramaic growhs during he las decade wih he proliferaion of IEEE 82.11 devices [1]. WLAN coninues o be he dominaing wireless infrasrucure for wireless consumer elecronic (CE) devices wih various applicaions [2], [3], [4]. IEEE 82.11 WLAN Task Groups are pursing gigabi wireless communicaions o furher increase he hroughpu and Manuscrip received March 29, 217; acceped June 16, 217. Dae of publicaion July 2, 217. 1 This work was suppored in par by Wireless@ Virginia Tech. (Corresponding auhor: Xiaofu Ma.) Xiaofu Ma is wih he Virginia Polyechnic Insiue and Sae Universiy, Blacksburg, VA 246 USA (e-mail: xfma@v.edu). Qinghai Gao is wih Qualcomm Aheros, San Jose, CA 9511 USA (e-mail: qinghai.gao@gmail.com). Ji Wang is wih he Virginia Polyechnic Insiue and Sae Universiy, Blacksburg, VA 246 USA (e-mail: raceyw@v.edu). Vuk Marojevic is wih he Virginia Polyechnic Insiue and Sae Universiy, Blacksburg, VA 246 USA (e-mail: maroje@v.edu). Jeffrey H. Reed is wih he Virginia Polyechnic Insiue and Sae Universiy, Blacksburg, VA 246 USA (e-mail: reedjh@v.edu). specral efficiency for he growing number of CE devices [5]. A ypical muliple anenna WLAN sysem includes one access poin (AP) and several mobile saions. APs have been MIMO capable since he release of IEEE 82.11n [6]. The new IEEE 82.11 sandards, IEEE 82.11ac [7] and 82.11ax [8], add muli-user MIMO (MU-MIMO) as he key echnology o improve sysem hroughpu. In paricular, an MU-MIMO sysem enables a MIMO capable AP o communicae wih muliple users simulaneously hrough spaial diversiy exploiaion and spaial muliplexing. Compared wih he radiional MIMO sysem, he use of MU- MIMO sysems reduces he requiremens of he end user devices, which do no need o have muliple anennas. However, he deploymen of MU-MIMO sysems faces several challenges. This paper addresses one such challenge relaed o acquiring accurae channel sae informaion (CSI) ha describes he properies of he communicaion link. Inaccurae CSI can mislead he beam seering a he ransmier and herefore resuls in severe cross-user inerference a he receiver, causing significan hroughpu degradaion. Explici channel sounding is necessary o acquire CSI in an MU-MIMO capable WLAN sysem. During one explici channel sounding process, a pre-known sequence of daa is ransmied o he receivers on he arge channel. Based on he received signal, he receiver esimaes he CSI and sends his informaion back o he ransmier. The CSI accuracy depends no only on he correcness of he esimaion process, bu also on he sounding updae frequency. However, he increase of channel sounding frequency reduces he daa ransmission windows. This creaes a rade-off beween collecing up-o-dae CSI and maximizing he ransmission ime. Deriving a suiable sounding inerval for an MU-MIMO sysem in a real scenario is challenging because of he dynamic and unpredicable channel condiions. For example, in he channel environmens wih saic devices and few human aciviies, he channel says relaively sable, which requires less frequen channel sounding. In conras, device and environmen mobiliy can creae a fas changing channel environmen, which would dramaically degrade he sysem hroughpu unless he CSI is frequenly updaed. Therefore, he sounding-ransmission rade-off needs o be coninuously evaluaed.
2 This paper exends he sae of he ar in channel sounding for MU-MIMO sysems by providing a heoreical analysis and experimenal resuls based on exensive over-he-air channel measuremens in an indoor WLAN environmen. The resuls show how dynamic sounding can provide beer hroughpu over saic sounding a minimal compuaional and signaling overhead. The main conribuions are summarized as follows. (1) The selecion of he channel sounding frequency is characerized and evaluaed for single user ransmi beamforming (SU-TxBF) and MU-MIMO in WLAN. The mahemaical formaion is used o formally describe he problem and explain he performance rend. Raher han solving he problem analyically, his paper presens a pracical approach: Exensive channel measuremens were conduced in boh, saic and dynamic indoor scenarios using MU-MIMO es nodes o demonsrae he exisence of a channel-specific opimal sounding inerval. (2) A dynamic sounding approach of low complexiy is derived o improve he effecive hroughpu in real WLAN indoor environmens. Using a developed 82.11ac emulaor seeded wih measured channel informaion, he proposed approach is shown o achieve hroughpu improvemens of up o 31.8%. The res of he paper is organized as follows: Secion II presens he problem and relaed work. In Secion III, he rade-off beween sounding overhead and hroughpu is evaluaed. A pracical dynamic sounding approach is proposed and evaluaed in Secion IV. Secion V concludes he paper. II. PROBLEM STATEMENT AND RELATED WORK A. Problem Saemen MU-MIMO is a promising echnology for increasing he sysem hroughpu of WLAN sysems. A suiable channel sounding sraegy is criical for his. Too infrequen sounding operaions cause more cross-user inerference because of he ou-daed CSI, whereas oo frequen sounding operaions unnecessarily ake up he effecive ransmission ime, boh leading o performance deficis. The problem hus consiss in finding a suiable operaional poin ha rades he CSI accuracy for effecive ransmission ime as a funcion of he radio environmen. In he conex of IEEE 82.11ac and fuure WLAN sysems, his paper assesses he following wo aspecs relaed o explici channel sounding: (1) Wha is he opimal sounding inerval ha maximizes sysem hroughpu? (2) How o dynamically adap he sounding inerval for rapidly changing radio environmens? B. Background and Relaed Work To increase he MU-MIMO sysem hroughpu, many research effors focus on he sraegy of user grouping. These works assume ha perfec CSI informaion is available a he ransmier. The scheme presened by Dimic and Sidiropoulos [9] greedily groups users based on he esimaed capaciy for improved sysem hroughpu. Channel orhogonaliy can also be an effecive meric o group users. For example, a heurisic approach was proposed by Yoo and Goldsmih [1] o gaher he mos orhogonal users in one group and disperse highlycorrelaed users over differen ime slos. The use of group membership and group idenifiers for managing MU-MIMO downlink ransmissions [11] was sudied specifically for WLAN sysems. The media access conrol (MAC) proocol design is anoher research area for MU-MIMO sysems. A modified CSMA/CA proocol [12] for MU-MIMO sysems was proposed, focusing on ACK-replying mechanisms o improve sysem hroughpu, where CSI is assumed o be known for he minimum mean square error (MMSE) precoding. The proporional fair allocaion mechanism for MU-MIMO ransmission was adoped by Valls and Leih [13] and assumes ha he AP has full knowledge of he channel. These works have paved he pah for pracical implemenaions of MU-MIMO in WLAN sysems; however, hey do no evaluae he channel sounding overhead and CSI imperfecion, which direcly affec MU- MIMO hroughpu. Some perspecives of CSI have been analysed for MU- MIMO in commercial wireless sysems using numerical evaluaion. Training and scheduling aspecs of MU-MIMO schemes [14] ha rely on he use of oudaed CSI were invesigaed, showing ha cerain MU-MIMO hroughpu can be achieved even wih fully oudaed CSI. For he downlink of MU-MIMO-based FDD sysems, he channel feedback mechanism [15] was sudied o consider user diversiy and he channel correlaions in boh ime and frequency. Recen lieraure also provides experimenal resuls for MU- MIMO. Balan e al. [16] demonsraed ha he sysem daa rae of he MU-MIMO sysem using channel feedback was much beer han wihou channel feedback. The impac of CSI compression on he feedback overhead was invesigaed by Xie e al. [17]. In order o reduce he sounding overhead, implici sounding [18] may ouperform explici sounding wih lower ime overhead. However, implici sounding requires more compuaion for boh he channel and ransceiver radio frequency (RF) chain calibraion o mainain full channel reciprociy. In addiion, imperfec CSI a he ransmier degrades he hroughpu of a MU-MIMO sysem hroughpu more severely han ha of a basic MIMO sysem. The sandard form of channel sounding in IEEE 82.11ac is only explici [19], which requires he use of channel measuremen frames. The non-negligible overhead produced by explici channel sounding moivaes research on finding he opimal sounding frequency. The relaionship beween sounding frequency and MU-MIMO sysem hroughpu needs boh heoreical and experimenal invesigaions. There are heoreical conribuions for deriving he opimal sounding inerval for a basic MIMO sysem. Zhang e al. [2], for insance, assumed he Rayleigh block-fading channel. These soluions canno be direcly applied o he MU-MIMO sysem. In addiion, in pracical deploymens where environmenal or device mobiliy can cause significan channel variaions, a predefined
3 sounding inerval would no be suiable. The curren lieraure lacks sysem-level analyses and evaluaions ha consider dynamic sounding frequency as anoher degree of freedom for pracical WLAN sysems. This paper analyzes and experimenally validaes he imporance of dynamic channel sounding for emerging and fuure WLAN deploymens. III. STATIC OPTIMAL SOUNDING INTERVAL A well-chosen sounding inerval maximizes he effecive specral efficiency, which is defined as he oal daa rae delivered o all saions over he ransmission bandwidh. This effecive specral efficiency depends on wo major facors: he effecive ransmission percenage and he insananeous daa rae as a funcion of ime. This secion uses he effecive specral efficiency o analyze he performance of SU-TxBF and MU-MIMO, boh being par of he new generaion WLAN sysems. The goals of his secion are (1) o show he exisence of a channel-specific sounding inerval using he over-he-air measuremens, and (2) o demonsrae he dependency of his saic opimal sounding inerval on he differen sysem parameers. The saic opimal sounding inerval is he one among all possible consan sounding inervals ha leads o he highes effecive specral efficiency. The channel model is presened firs, followed by he opimizaion problem formulaion and evaluaion using an IEEE 82.11ac emulaor seeded wih real channel measuremens colleced for differen indoor scenarios. A. Channel Model In his paper, he impulse response of he channel a ime is denoed by h,, where is he delay variable. Due o he mulipah fading effec, h, can be expressed as, h N k k1 where N represens he number of aps, a δ τ τ (1) k ak represens he weigh of he ap k and k represens he delay on he k h ap. The impulse responses are based on he cluser model [18], and he power of delayed responses decays linearly on a log-scale, i.e., he power decays exponenially. The powers of he aps in overlapping clusers are summed a each delay. The frequency domain represenaion is adoped for MIMO operaion analysis. The sysem analysis is presened for only one single subcarrier for noaion simpliciy, bu he analysis applies o any number of subcarriers. B. Effecive Specral Efficiency The ime consumed for each daa ransmission period is denoed by T and includes wo pars: channel sounding and daa ransmission. T can also be viewed as he ime gap beween wo successive sounding operaions. The operaion ime for a sounding operaion is denoed by T S. The value of TS differs for differen number of user(s) served by he SU- TxBF or MU-MIMO ransmission. The effecive ransmission ime percenage hen becomes ( 1 TS / T ). I is assumed ha he channel undergoes block fading and each block lass for a very small ime. Thus, he effecive specral efficiency Ce ( T ) during T can be formulaed as M TS 1 Ce ( T ) (1 ) CT ( m ). (2) S T M m1 Here, M T /, and C () is he specral efficiency. C () is measured as he expeced value over channel Nss realizaion and ransmission of 1 i 1 log 2 i, where is he number of ransmission sreams and T i Nss represens he SNR of he i h spaial sream. I is obvious from (2) ha ( 1 T / ), he firs par of C ( T ), S increases wih increasing T. The value of he second par of M 1 he equaion, CT ( m ), also changes wih T S. M m1 The problem formulaion for he opimal sounding inerval can herefore be wrien as max Ce ( T ) s.. T T. S (3) In order o maximize he effecive specral efficiency wih respec o he channel sounding inerval T, i is necessary o T. T analyze he impac of T on i The number of ransmi anennas is denoed by N. For SU- TxBF using zero-forcing precoding, he SINR of he spaial sream i afer from he channel esimaion ime can be expressed as (wihou losing generaliy, is se o ) i 2 w s i h j w s i s i P h i s i N P ji 2 e, (4) where P i is he power allocaed o spaial sream i, hs i is he 1-by- N frequency-domain channel response vecor of spaial sream i a ime, w si is he N -by-1 sreering vecor for spaial sream i based on he channel esimaion a ime, and N is he channel noise power. The expression P ji j s i s i h w is he erm of cross-sream 2 inerference for sream i a he receiver. In he MU-MIMO case, when zero-forcing precoding is used for muli-sream seering, he SINR of he spaial sream i for user n afer from he channel esimaion ime is
4 ni, w ni, hni, ni, P 2 2 n, j h w h w ni, n, j n, j ni, m, j ji mn N P P where Pni, is he power allocaed o sream i for user n, h ni, is he 1-by 2, (5) N frequency-domain channel response vecor of spaial sream i for user n a ime, and he wni, is N -by-1 sreering vecor for spaial sream i a user n based on he channel esimaion a ime. The expression P ji n, j ni, n, j 2 h w in (5) is he cross-sream mn n, j ni, m, j inerference a user n, and P 2 h w is he cross-user inerference. The opimal soluion of he formulaion in (3) would maximize he effecive specrum efficiency. However, his problem is very complex o solve for a given channel and is impracical as he channel saus coninuously changes in a real environmens. I is herefore infeasible o derive he closed-form soluion for IEEE 82.11ac consumer elecronic sysems. Insead, he nex subsecion evaluaes he ransmission rade-off using channel measuremens in differen indoor environmens as a pracical approach o he problem. C. Evaluaion of Trade-off Beween Accurae CSI and Sounding Overhead (1) Experimenal Sysem Seup In his secion, he dependency of he opimal sounding inerval on he differen sysem parameers is evaluaed in a WLAN-based MU-MIMO environmen. Consider a ypical WLAN opology wih one AP and 12 saions. The AP is equipped wih four anennas and each saion is equipped wih a single anenna. The sysem can serve up o hree saions simulaneously on he downlink, which is ypical for a pracical WLAN scenario. The measuremens of he acual channel informaion were conduced using four es nodes in an office environmen as illusraed in Fig. 1. Each es node was equipped wih four anennas and a 4 4 four-sream 82.11ac ransceiver chipse. The over-he-air ransmissions were conduced over a 4 MHz channel. One es node operaed as he ransmier AP, and he remaining hree operaed as he receiver saions. The 12 receive anennas were spaced apar from one anoher and disribued randomly in he office. Those 12 ses of channel samples are used o mimic he channel from one four-anenna AP o 12 single-anenna saions. The channel measuremens were colleced under wo scenarios, he low Doppler and high Doppler scenarios. The low Doppler scenario refers o a low-mobiliy radio environmen wih no human aciviy or device movemen during he channel measuremens. Channel samples under he high Doppler scenario were colleced when significan human movemen exised. Since he anenna movemen and human movemen are relaive, hese channel samples were used o evaluae he high dynamic environmen. Noice ha here are sill random variaions in boh he low and he high Doppler scenarios. Fig. 1 Measuremen seup skech map Wih he measured channel informaion, rials wih differen sounding inervals can be conduced using idenical channels for comparison. A measuremen-driven MU-MIMO-OFDM emulaor was rigorously implemened according o he IEEE 82.11ac specificaions. This emulaor was seeded wih he over-he-air channel informaion o es he performance of he MAC and physical (PHY) layer algorihms. Singular value decomposiion (SVD) was employed as he precoding scheme, and he MMSE receiver was implemened o miigae he cross-user inerference. Since he focus of his paper is he selecion of he sounding inerval, oher possible effecs on he sysem performance have been carefully eliminaed. In paricular, i is assumed ha he AP adaps he opimal modulaion and coding scheme (MCS) during each simulaion rial. The evaluaion parameers of he IEEE 82.11ac sysem are summarized in Table Ι. Parameers TABLE I EVALUATION PARAMETERS Values Maximum AMPDU Duraion 2 ms MPDU Lengh 1556 bis MSDU Lengh 158 bis SIFS Duraion.16 ms Bandwidh 4 MHz Number of OFDM Subcarriers for Daa 18 Guard Inerval for OFDM Symbol 4 ns In a single user ransmission scenario, he sounding overhead of 82.11ac comes from he explici feedback mechanism, i.e., he informaion exchange beween he AP and he mobile saions. This informaion exchange during each channel sounding process is as follows: The AP firs broadcass a Null Daa Packe Announcemen (NDPA). Afer a Shor Inerframe Space (SIFS) ime inerval, he AP hen
5 sends ou a Null Daa Packe (NDP) o sound he ransmission channel. From he saion s perspecive, afer receiving he NDP from he AP, i wais for a SIFS inerval and responds o he AP wih a compressed beamforming (CBF) repor. Hence, he ime required for each channel sounding operaion for single user ransmi beamforming can be expressed as T T T T T T T, (6) S _ SU NDPA SIFS NDP SIFS CBF SIFS where T NDPA, T SIFS, T and T are he ime duraions for he NDP CBF NDPA ransmission, SIFS, NDP ransmission and CBF ransmission, respecively. Noice ha NDPA and NDP are used by he AP o conrol he sounding inerval which does no need o be known by he saions. The sounding duraion for MU-MIMO in WLAN differs from ha of he single user case. I also includes he beamforming repor poll(s) for all saions as well as a sequence of he saions CBF repors. The sounding procedure for MU-MIMO iniiaes exacly he same as he SU-TxBF sounding procedure. The AP firs broadcass an NDPA. Afer a SIFS ime inerval, he AP sends ou an NDP o sound he ransmission channel. However, o rerieve he channel informaion from each STA, he MU sounding procedure needs one or more Beamforming Repor Poll (BRP) frames o collec responses from all he saions, one by one. A poll frame is no needed for he firs saion, which is specified in he NDPA frame, bu saring from he second and subsequen saions, BRP frame for each saion is required. The AP will inegrae all he received CBF repors ogeher ino a seering marix. Therefore, he ime required for one channel sounding operaion under a MU-MIMO ransmi beamforming can be derived from (6) and expressed as T T T T T S _ MU NDPA SIFS NDP SIFS 1 N T T N T T u SIFS CBF u SIFS BRP, (7) where N represens he number of users in he muli-user u ransmission group, and T is he ime duraion for a BRP BRP frame ransmission. (2) Evaluaion Resuls Fig. 2 plos he oupu SINR as a funcion of he sounding inerval for SU-TXBF ransmission, MU-MIMO ransmission wih 2 users (MU2), and MU-MIMO ransmission wih 3 users (MU3). More precisely, he SINR for SU, MU2, and MU3 cases are compared wih sounding inervals ranging from 2 ms o 4 ms in low (Fig. 2a) and high (Fig. 2b) Doppler scenarios. The SINR for SU-TxBF remains nearly consan. As opposed o MU-MIMO, SU-TxBF does no need o deal wih iner-user inerference. As a resul, he channel esimaion is demonsraed here o be fresh enough even for a 4 ms sounding gap for SU-TxBF. The SINR would evenually degrade when he gap increases furher. MU-MIMO is more sensiive o inaccurae CSI because he spaial sream seering is done in such a way ha he iner-user inerference is minimized for MU ransmission as opposed o maximizing he user s SINR for SU-TxBF. Fig. 2 correspondingly shows how he SINR value for MU-MIMO ransmission quickly decreases wih he sounding gap even in he low Doppler scenario (Fig. 2a) and more quickly for he high Doppler scenario (Fig. 2b). This is because he CSI accuracy for beam seering degrades wih increasing ime gap beween he channel esimaion and he daa ransmission. I is also observed ha he SINR per spaial sream decreases wih he increasing number of users in he MU ransmission in he sysem. For example, afer 2 ms since he las channel sounding operaion, he SINR has dropped by 4 db (from 31 o 27 db) for MU2 and by 5 db (from 27 o 22 db) for MU3 in he low Doppler scenario. The SINR degradaion is much more severe for he high Doppler scenario: A 2 ms, for MU2 and MU3, he SINR values are 14 and 8 db. These resuls illusraes he significan impac of fas changing channels on SINR for MU ransmission and he relaive robusness for SU-TxBF. SINR (db) 35 3 25 2 15 1 5 SU-TxBF MU-2 users MU-3 users 1 2 3 4 Sounding Gap Time (ms) (a) SINR (db) 35 3 25 2 15 1 5 SU-TxBF MU-2 users MU-3 users 1 2 3 4 Sounding Gap Time (ms) (b) Fig. 2. SINR as a funcion of sounding inerval in (a) he low Doppler scenario and (b) he high Doppler scenario. Fig. 3 shows he corresponding oal PHY layer hroughpu as a funcion of he sounding gap. For boh he low (Fig. 3a) and high (Fig. 3b) Doppler scenario, he oal PHY layer hroughpu says relaively consan for SU-TxFB. This is because he sufficienly high SINR for SU-TxFB can suppor he highes MCS, which leads o consan hroughpu. This confirms o he SINR resuls shown in Fig. 2a and Fig. 2b. On he oher hand, he hroughpu decreases monoonically for MU2 and MU3 ransmissions. This illusraes he impac of cross-user inerference due o he oudaed CSI on sysem performance. The drop of he sysem hroughpu under he same sounding gap is enlarged by he number of users wihin MU ransmission. Specifically for he low Doppler scenario, when he sounding gap increases from 2 o 5 ms, he PHY hroughpu of MU3 drops by 18% compared o only 5.6% drop for MU2. This is because he increased number of users wihin an MU ransmission worsens he cross-user inerference. The same phenomenon can be also observed from he high Doppler scenario.
6 Toal PHY Throughpu (Mbps) 5 45 4 35 3 25 2 MU-3 users MU-2 users SU-TxBF Toal PHY Throughpu (Mbps) 5 45 4 35 3 25 2 MU-3 users MU-2 users SU-TxBF significan affec he hroughpu even when he sounding inerval becomes as large as 1 ms or more. This is differen for he high Doppler scenario (Fig. 4b), where he MAC hroughpu quickly drops when passing he opimal sounding inerval because of he iner-user inerference caused by he rapidly changing environmen and inaccurae CSI. The highes hroughpu for he high Doppler case is achieved for a 1-2 ms sounding inerval. 4 Low Doppler 4 High Doppler 15 15 35 35 1 1 2 3 4 Sounding Gap Time (ms) (a) 1 1 2 3 4 Sounding Gap Time (ms) Fig. 3. Toal PHY hroughpu as a funcion of sounding inerval in (a) he low Doppler scenario and (b) he high Doppler scenario. Compared wih he low Doppler scenario, he hroughpu drop in he high Doppler scenario as he sounding inerval increases becomes more significan as shown in Fig. 3b. When he sounding gap is less han 1 ms, he PHY layer hroughpu of MU3 ouperforms ha of MU2, and boh ouperform SU-TxBF. When he sounding gap approaches 5 ms, he PHY layer hroughpus of MU2 and MU3 drop o he same level as SU-TxBF. Beyond 5 ms, MU2 has a higher oal PHY layer hroughpu han MU3 and boh are below SU- TxBF hroughpu. I is also observed from Fig. 3 ha for he same lengh of sounding gap, a hroughpu difference exiss beween he low and high Doppler scenarios, and his difference is more significan when he number of users wihin MU ransmission is larger. For example, when he sounding gap is 5 ms, he hroughpu difference for MU2 beween high and low Doppler scenarios is 165 Mbps, whereas he difference becomes 22 Mbps for MU3. When he sounding gap increases o 1 ms, he difference becomes 18 Mbps and 245 Mbps, respecively. When he sounding overhead is considered, he hroughpu is he acual hroughpu ha can be achieved and delivered o he end users. This is called effecive hroughpu or MAC hroughpu. Fig. 4 shows he effecive hroughpu as a funcion of he sounding gap for MU3. All curves illusrae a similar rend: when he sounding gap is small, he sounding overhead dominaes he air ime, which causes he low effecive hroughpu. When he sounding gap increases, he relaive sounding overhead correspondingly decreases and he hroughpu increases unil a poin when he oudaed CSI dominaes. Noe ha he PHY layer hroughpu, which does no accoun for he sounding overhead, decreases over ime since he las sounding operaion because of he decreasing accuracy of he available CSI over ime. The opimal sounding inerval is around 5 ms for he low Doppler scenario (Fig. 4a). Since he channel environmen is relaively sable, he oudaed channel informaion does no (b) MAC Throughpu (Mbps) 3 25 2 15 1 5 Pah Loss 45dB Pah Loss 9dB 5 1 15 2 Sounding Gap (ms) MAC Throughpu (Mbps) 3 25 2 15 1 5 Pah Loss 45dB Pah Loss 9dB 5 1 15 2 Sounding Gap (ms) (a) (b) Fig. 4. Effecive hroughpu (MAC hroughpu) as a funcion of he sounding inerval under differen pah losses for (a) he low Doppler scenario and (b) he high Doppler scenario. I is observed from Fig. 4 ha he opimal sounding inerval varies no only under differen channel variaion condiions, bu also under differen insananeous pah losses. In paricular, he effecive hroughou degrades as he pass loss increases in eiher low or high Doppler scenario. This is reasonable as larger pass loss leads o lower SINR. In he low Doppler scenario, he opimal sounding inerval is 5 ms if he pah loss is 45 db, whereas he opimal value becomes around 1 ms if he pah loss is 9 db. A similar phenomenon can be observed in he high Doppler scenario, where he opimal sounding inerval under 45 db pass loss is around 1 ms, which is lower han ha under 9 db pass loss (around 2ms). Through he over-he-air measuremens, his secion shows he exisence of he channel-specific opimal sounding inerval and is dependency on he channel variaion and pah losses. The channel variaion indicaes he device or environmen moion, whereas he pah loss reflecs he geolocaion of boh he AP and he saion. The SINR and Doppler profile of each saion s spaial sream may change rapidly based on he insananeous channel environmen. Since he moion in he environmen and he geolocaion of he saions are dynamic and unpredicable, i is necessary o design an algorihm ha dynamically deermines a suiable operaional poin ha rades CSI accuracy for effecive ransmission ime in real ime as a funcion of he radio environmen.
7 IV. SOUNDING IN DYNAMIC ENVIRONMENTS So far, he deerminaion of he opimal sounding inerval has been invesigaed in a radio environmen characerized by a low or high Doppler scenario. In dynamic indoor environmens, he channel condiions change unpredicably because of he random movemen of human and wireless devices, or oher changes o he environmen. The sounding inerval requires dynamic adapaion based on he insananeous channel condiions. Thus, a dynamic sounding approach is designed ha (1) is aware of he insananeous channel condiion, (2) infers he suiable sounding inerval, and (3) adaps he sounding inerval. This is enabled by coninuously esimaing he saisics of he insananeous channel condiions, calculaing a reference hroughpu, and providing inernal feedback when he reference hroughpu drops. The reference hroughpu indicaes he accumulaed MAC layer hroughpu since he las sounding operaion. A. Proposed Dynamic Sounding Approach To race he ime variaion of he radio environmen, a Doppler specrum profile can be mainained from he feedback CSI o esimae he opimal sounding inerval. However, boh he Doppler specrum profile creaion and he heoreical calculaion of he opimal sounding inerval require a lo of compuing power and ime, which is very challenging for he real-ime sysem implemenaions. To simplify he sysem design, a sounding adjusmen is designed such ha i depends on he accumulaed daa rae, which can be easily colleced a he AP. Boh he sounding overhead and he insananeous ransmission condiions are considered o esimae he ime correlaion of he channel variaion and o calculae he effecive hroughpu. The accumulaed effecive hroughpu is used as he reference hroughpu o adjus he sounding inerval. All 82.11ac daa frames are sen in an aggregaed MAC proocol daa uni (AMPDU). The reference hroughpu is calculaed afer he n h AMPDU ransmission since he las sounding operaion using R TH n s n N u m1 j1 n m1 AMPDU D m, j T T m, (8) which represens he raio beween he esimae of he accumulaed successfully-ransmied AMPDUs during he downlink MU ransmission and he ime period. This period includes he overhead of he las sounding operaion and he ime for accumulaed downlink AMPDU ransmissions. Dm, jsands for he successfully ransmied daa amoun for he m h AMPDU of user j since he las sounding operaion, and TAMPDU m is he ime duraion of he m h downlink AMPDU ransmission. The basic idea of he dynamic sounding approach is o rigger he sounding operaion when he reference hroughpu sars degrading. In he proposed dynamic sounding approach, which is formally described in Algorihm 1, he AP regularly esimaes he reference hroughpu as a funcion of ime since he las sounding operaion. Afer a sounding operaion, he PHY layer hroughpu is expeced o improve and hen degrade over ime unil he nex sounding operaion. The reference hroughpu, on he oher hand, iniially increases as he sounding overhead dominaes he effecive hroughpu. However, he reference hroughpu sars dropping as he increasing CSI inaccuracy over ime becomes he dominan facor degrading he effecive hroughpu, which coninues degrading unil here is anoher sounding operaion. Since he reference hroughpu does no reflec he insananeous channel saisic, bu, raher, an accumulaed value represening a ime inerval, he impac of he insananeous flucuaion of noise on riggering a sounding operaion is miigaed. The proposed approach also works when uplink ransmission exises, since he reference hroughpu is a meric ha reflecs he hroughpu rend of he downlink ransmission over ime. Algorihm 1 provides he deails of he proposed dynamic sounding approach. From (8), i can be seen ha he algorihm complexiy is low and, hence, well suied for commercial MU-MIMO sysems. Algorihm 1. Dynamic Sounding Approach Sage 1: Iniializaion 1 Flag for wheher sounding is needed: s 1. 2 Flag for wheher he curren AMPDU packe is he firs one afer sounding: f. 3 The ransmied AMPDU index: n. Sage 2: MU Downlink Operaion 4 while MU Downlink Transmission 5 if s 1 6 Operae channel sounding. 7 f 1. 8 end if 9 Transmi AMPDU using rae adapaion algorihm. 1 n n 1. 11 Updae R n using (8) TH 12 if f 1 or R n 1 R n TH 13 Se s and prepare for sending he nex AMPDU. 14 f. 15 else 16 Se s 1and prepare for sounding. 17 end if 18 end while B. Evaluaion Consider a WLAN sysem wih one AP and 12 saions. The 82.11ac emulaor discussed in Secion III wih he overhe-air channel measuremens is used as he evaluaion plaform. Since he sysem can suppor up o hree saions for TH
8 MU-MIMO ransmission, i is considered ha any hree saions ou of he 12 are grouped. The performance of he proposed approach is evaluaed using he colleced channel measuremens, including he low and high Doppler scenarios. The dynamic scenario is mimicked by alernaing beween he low and high Doppler channel samples. Specifically, he channel samples are rearranged ino an alernaion beween high and low Doppler condiions wih alernaing 5ms of high and low Doppler channel samples. As discussed in Secion II, conribuions exis in lieraure ha derive he opimal saic sounding inerval for a basic MIMO sysem assuming he Rayleigh block-fading channel, such as [2]. These soluions are derived for a basic MIMO sysem, and canno be direcly applied o MU-MIMO sysems. In addiion, in a pracical scenario where environmenal or device mobiliy can cause significan and unpredicable channel variaions, where he pre-compued sounding inerval may no be efficien. In order o experimenally validae he imporance of dynamic channel sounding for emerging and fuure WLAN deploymens, we compare he proposed approach wih wo saic schemes ha correspond o he sae-of-he-ar soluions presened in [2]. These wo schemes one has a shor and he oher a long sounding inerval provide he opimal fixed sounding inervals for he low and high Doppler scenarios. We call hem low Doppler approach (LDA) and high Doppler approach (HDA). Noice ha he channel variaion level canno be prediced in a real radio environmen, so i is no feasible o know he opimal fixed sounding inerval. Fig. 5 illusraes he sounding evens for he hree approaches and Fig. 6 compares he performance. Fig. 5 indicaes he ime samps for channel sounding as riggered by he AP using he hree schemes in an example 4 ms period wih alernaing high-low Doppler condiions. As expeced, under boh condiions, he sounding inerval under HDA is always around 11ms and he sounding inerval under LDA is around 43 ms. In conras, he sounding inerval of he dynamic sounding approach varies according o he change of he radio environmen, which illusraes is channel awareness funcionaliy. Paricularly, he average sounding inerval of he proposed approach in he high Doppler scenario is around 11 ms, which increases o around 41 ms for he low Doppler case. As can be observed from Fig. 5, he operaion aciviy of he dynamic sounding approach says almos he same as ha of LDA in low Doppler scenario and ha of HDA in he high Doppler scenario. This illusraes he accuracy of he channel awareness process of he proposed dynamic sounding approach. The hroughpu improvemen of approach X over approach Y is defined as (R X R Y ) / R Y, where R X and R Y are he hroughpus achieved by approach X and Y. Fig. 6 shows he hroughpu improvemen of he proposed dynamic sounding approach over LDA (green bars) and over HDA (yellow bars) in he high Doppler, low Doppler, and alernaing condiions. Low-Doppler Sounding Approach High-Doppler Sounding Approach Dynamic Sounding Approach Channel Sounding Even Record 5 1 15 2 25 3 35 4 ime (ms) 5 1 15 2 25 3 35 4 ime (ms) 5 1 15 2 25 3 35 4 ime (ms) Fig. 5. Example sounding even records for he wo opimal saic sounding approaches, LDA (op) and HDA (middle), and he proposed dynamic sounding approach (boom). Grey shaded areas indicae he high Doppler scenario and whie areas he low Doppler scenario. The black solid lines illusrae he sounding evens. Throughpu Improvemen (%) 4 35 3 25 2 15 1 5-5 Throughpu improvemen of Dynamic Sounding vs. LDA Throughpu improvemen of Dynamic Sounding vs. HDA High Doppler Low Doppler Alernaion Fig. 6. Throughpu improvemen of dynamic sounding over LDA and HDA. I is observed ha in he high Doppler scenario, he hroughpu improvemen of dynamic sounding over LDA is as high as 31.8%. This illusraes he benefi of operaing frequen sounding in rapidly changing radio environmens. In addiion, he hroughpu improvemen of he proposed dynamic sounding approach over HDA is 8.6%, which means ha he dynamic sounding approach ouperforms he scheme ha uses he opimal fixed-sounding inerval. This is because in he high Doppler scenario he channel variaion levels change over ime, and dynamic sounding approach can capure he channel changes imely o rigger a sounding operaions. Similarly, i is observed ha here is a 14.3% improvemen of using dynamic sounding approach over HDA in he low Doppler scenario. Dynamic sounding and LDA achieve similar performance, where he difference is 3.1%. In he alernaion condiions, however, none of he fixed approaches achieve he same level of performance as he proposed
9 soluion. In paricular, dynamic sounding ouperforms saic sounding by 19.8% (LDA) and 1.9% (HDA). The proposed approach explores he change in he channel environmen and rades he CSI accuracy for effecive ransmission ime as he funcion of observed channel variaions. V. CONCLUSIONS This paper has presened a framework for evaluaing channel sounding inervals wih pracical applicaions for MU-MIMO in WLAN. Using colleced channel measuremens, he rade-off beween CSI accuracy and he effecive ransmission ime has been evaluaed. and The opimal sounding inerval is analysed for he given channel environmen. For pracical scenarios where he indoor radio environmen can change unpredicably, a low-complexiy dynamic sounding approach is proposed ha updaes he sounding inerval o improve he effecive hroughpu. Reference hroughpu is seleced as he meric o deermine he sounding inerval. Under he colleced channel measuremens, i is shown ha significan hroughpu improvemens (up o 31.8%) for IEEE 82.11ac sysems can be achieved by using he proposed approach, especially in he highly dynamic environmens. This work can be exended o analyze oher commercial MU-MIMO / massive MIMO sysems and derive effecive channel sounding soluions for new radio bands in heerogeneous radio environmens. REFERENCES [1] J. Kim and I. Lee, "82.11 WLAN: hisory and new enabling MIMO echniques for nex generaion sandards," IEEE Commun. Mag., vol. 53, pp. 134-14, 215. [2] W. Yoo, Y. Jung, M. Y. Kim, and S. Lee, "A pipelined 8-bi sof decision vierbi decoder for IEEE82.11ac WLAN sysems," IEEE Trans. Consumer Elecron., vol. 58, pp. 1162-1168, 212. [3] H. Kasai, "Time-slo based even-driven nework swich conrol for informaion sharing in muliple WLANs," IEEE Trans. Consumer Elecron., vol. 59, pp. 521-529, 213. [4] F. Birlik, O. Gurbuz, and O. Ercein, "IPTVhome neworking via 82.11 wireless mesh neworks: an implemenaion experience," IEEE Trans. Consumer Elecron., vol. 55, pp. 1192-1199, 29. [5] V. Jones and H. Sampah, "Emerging echnologies for WLAN," IEEE Commun. Mag., vol. 53, pp. 141-149, 215. [6] E. Perahia, "IEEE 82.11n developmen: hisory, process, and echnology," IEEE Commun. Mag., vol. 46, pp. 48-55, 28. [7] R. Nee, "Breaking he gigabi-per-second barrier wih 82.11ac," IEEE Wireless Commun., vol. 18, pp. 4-4, 211. [8] D. Deng, K. Chen, and R. Cheng, "IEEE 82.11ax: Nex generaion wireless local area neworks," in 1h Inernaional Conference on Heerogeneous Neworking for Qualiy, Reliabiliy, Securiy and Robusness, Rhodes, Greece, 214, pp. 77-82. [9] G. Dimic and N. D. Sidiropoulos, "On downlink beamforming wih greedy user selecion: performance analysis and a simple new algorihm," IEEE Trans. Signal Process., vol. 53, pp. 3857-3868, 25. [1] T. Yoo and A. Goldsmih, "On he opimaliy of mulianenna broadcas scheduling using zero-forcing beamforming," IEEE J. Sel. Areas Commun., vol. 24, pp. 528-541, 26. [11] O. Aboul-Magd, U. Kwon, Y. Kim, and C. Zhu, "Managing downlink muli-user MIMO ransmission using group membership," in IEEE Consumer Communicaions and Neworking Conference, Las Vegas, USA, 213, pp. 37-375. [12] M. X. Gong, E. Perahia, R. Sacey, R. Wan, and S. Mao, "A CSMA/CA MAC proocol for muli-user MIMO wireless LANs," in IEEE Global Telecommunicaions Conference, Miami, USA, 21, pp. 1-6. [13] V. Valls and D. J. Leih, "Proporional fair MU-MIMO in 82.11 WLANs," IEEE Wireless Commun. Le., vol. 3, pp. 221-224, 214. [14] A. Adhikary, H. C. Papadopoulos, S. A. Ramprashad, and G. Caire, "Muli-user MIMO wih oudaed CSI: raining, feedback and scheduling," in 49h Annual Alleron Conference on Communicaion, Conrol, and Compuing, Monicello, USA, 211, pp. 886-893. [15] N. Jindal and S. Ramprashad, "Opimizing CSI feedback for MU- MIMO: radeoffs in channel correlaion, user diversiy and MU-MIMO efficiency," in IEEE 73rd Vehicular Technology Conference, Budapes, Hungary, 211, pp. 1-5. [16] H. V. Balan, R. Rogalin, A. Michaloliakos, K. Psounis, and G. Caire, "Achieving high daa raes in a disribued MIMO sysem," in Proc. of he 18h Annual Inernaional Conference on Mobile Compuing and Neworking, Isanbul, Turkey, 212, pp. 41-52. [17] X. Xie, X. Zhang, and K. Sundaresan, "Adapive feedback compression for MIMO neworks," in Proc. of he 19h Annual Inernaional Conference on Mobile Compuing & Neworking, Miami, USA, 213, pp. 477-488. [18] E. Perahia and R. Sacey, Nex generaion wireless LANs: 82.11 n and 82.11 ac: Cambridge Universiy Press, 213. [19] M. Gas, 82.11 ac: A survival guide: O'Reilly Media, Inc., 213. [2] L. Zhang, L. Song, M. Ma, and B. Jiao, "On he Minimum Differenial Feedback for Time-Correlaed MIMO Rayleigh Block-Fading Channels," IEEE Trans. on Commun., vol. 6, pp. 411-42, 212. Xiaofu Ma (S 13) received he B.S. degree in elecronics from Norhwes Universiy, Xi'an, China, in 28, and he M.S. degree in compuer science from Tongji Universiy, Shanghai, in 211, and is currenly pursuing he Ph.D. degree in elecrical engineering a Virginia Polyechnic Insiue and Sae Universiy, Blacksburg, USA. His research ineress include nework proocol and opimizaion, cogniive radio neworks, specrum sharing, chipless RFID, wireless healhcare and wireless local area neworks. Qinghai Gao received his B.S. and M.S. from Xidian Universiy, Xi an, China, in 1999 and 22 respecively, and he Ph.D. degree from Arizona Sae Universiy in 28. He is currenly wih Qualcomm Aheros, San Jose, CA. His research ineress are in he areas of cross-layer opimizaion and cooperaive communicaions in wireless neworks. Ji Wang received he B.S. degree in Elecrical Engineering from Harbin Insiue of Technology, Harbin, China, in 29, and he M.S. degree in applied mahemaics and saisics wih elecrical engineering minor in 212 from Universiy of Minnesoa, and is currenly pursuing he Ph.D. degree in elecrical engineering a Virginia Polyechnic Insiue and Sae Universiy, Blacksburg, USA. Her research ineress include cogniive radio nework proocol design and analysis of such sysems and neworks using game heory.
1 Vuk Marojevic received he Dipl.-Ing. degree from he Universiy of Hannover in 23 and he Ph.D. degree from he Universia Poliècnica de Caalunya (UPC), Spain, in 29, boh in elecrical engineering. He is currenly wih he Wireless @ Virginia Tech research group. His research ineress include sofwaredefined and cogniive radio echnology, he long-erm evoluion (LTE), wireless esbeds, infrasrucure and waveforms for reliable high-capaciy neworks. Jeffrey H. Reed (F 5) received he B.S.E.E., M.S.E.E., and Ph.D. degrees from he Universiy of California Davis, Davis, CA, USA, in 1979, 198, and 1987, respecively. He is he founder of Wireless @ Virginia Tech, and served as is direcor unil 214. He is he Founding Faculy member of he Ted and Karyn Hume Cener for Naional Securiy and Technology and served as is inerim Direcor when founded in 21. He is cofounder of Cogniive Radio Technologies (CRT), a company commercializing of he cogniive radio echnologies; Allied Communicaions, a company developing echnologies for 5G sysems; and for Power Fingerprining, a company specializing in securiy for embedded sysems. Dr. Reed is a Disinguished Lecurer for he IEEE Vehicular Technology Sociey. In 212, Dr. Reed served on he Presiden s Council of Advisors of Science and Technology Working Group ha examines ways o ransiion federal specrum o allow commercial use and improve economic aciviy.