Interference Alignment and Cancellation

Transcription

1 Intefeence Alignment and Cancellation Shyamnath Gollakota, Samuel David Peli and Dina Katabi MIT CSAIL ABSTRACT The thoughut of existing MIMO LANs is limited by the numbe of antennas on the AP. This ae shows how to ovecome this limitation. It esents intefeence alignment and cancellation (IAC), a new aoach fo decoding concuent sende-eceive ais in MIMO netwoks. IAC synthesizes two signal ocessing techniques, intefeence alignment and intefeence cancellation, showing that the combination alies to scenaios whee neithe intefeence alignment no cancellation alies alone. We show analytically that IAC almost doubles the thoughut of MIMO LANs. We also imlement IAC in GNU-Radio, and exeimentally demonstate that fo x MIMO LANs, IAC inceases the aveage thoughut by.x on the downlink and x on the ulink. Categoies and Subject Descitos C.. [Comute Systems Oganization]: Comute-Communications Netwoks Geneal Tems Algoithms, Design, Pefomance, Theoy Keywods Intefeence Alignment, Intefeence Cancellation Intoduction Multi-inut multi-outut (MIMO) technology is emeging as the natual choice fo futue wieless LANs. The cuent design, howeve, meely elaces a single-antenna channel between a sende-eceive ai with a MIMO channel. The thoughut of such a design is always limited by the numbe of antennas e access oint (AP) [, 9]. Intuitively, if each node has two antennas, the client can simultaneously tansmit two ackets to the AP. The AP eceives a linea combination of the two tansmitted ackets, on each antenna, as shown in Fig.. Hence, the AP obtains two linea equations fo two unknown ackets, allowing it to decode. Tansmitting moe concuent ackets than the numbe of antennas on the AP simly inceases intefeence and events decoding. Thus, today the thoughut of all actical MIMO LANs is limited by the numbe of antennas e AP. This ae intoduces Intefeence Alignment and Cancellation (IAC), a actical scheme to ovecome the antennas-e-ap thoughut limit in MIMO LANs. IAC synthesizes two intefeence management techniques: intefeence alignment and intefeence cancellation, showing that the combination imoves efomance in scenaios whee neithe intefeence alignment no cancellation alies alone. To get a feel fo how IAC woks, conside again a -antenna client that uloads two concuent ackets to a -antenna AP. Say we have Pemission to make digital o had coies of all o at of this wok fo esonal o classoom use is ganted without fee ovided that coies ae not made o distibuted fo ofit o commecial advantage and that coies bea this notice and the full citation on the fist age. To coy othewise, to eublish, to ost on seves o to edistibute to lists, equies io secific emission and/o a fee. SIGCOMM 09, August 7, 009, Bacelona, Sain. Coyight 009 ACM /09/08 Client h h h h y = + h h = h h y + y AP y Figue : Thoughut of cuent MIMO LANs is limited by the numbe of antennas e AP. The h i j s ae known channel coefficients, and the i s ae concuent ackets. The client tansmits two concuent ackets. The AP eceives a diffeent linea combination of the tansmitted ackets on each antenna, which it solves to obtain the ackets. Clients APs Figue : IAC Examle. AP decodes acket and sends the decoded acket on the Ethenet to AP which then efoms intefeence cancellation to subtact. As a esult AP can decode and. a second x client-ap ai on the same wieless channel and within intefeence ange. Can the second client-ap ai concuently uload a thid acket? In existing MIMO LANs, the thee concuent ackets intefee. As a esult, each of the two APs gets two linea equations with thee unknown ackets, and hence cannot decode. In contast, IAC allows these thee concuent ackets to be decoded. To do so, IAC exloits two oeties of MIMO LANs: ) MIMO tansmittes can contol the alignment of thei signals at a eceive, and ) APs ae tyically connected to a backend Ethenet, which they can use fo coodination. Thus, in IAC, the two clients encode thei tansmissions in a secial way to align the second and the thid ackets at AP but not at AP, as shown in Fig.. As a esult, AP can teat the second and thid ackets as one unknown; i.e., AP has the equivalent of two equations with two unknowns, allowing it to decode the fist acket,. AP then sends the decoded acket on the Ethenet to AP, which can now efom intefeence cancellation to subtact the effect of the known acket. As a esult, AP is left with two linea equations ove two unknown ackets, and, which it can decode. The system delives thee ackets e time unit. Hence, its thoughut is not bounded by the numbe of antennas e AP. Note the synegy between intefeence alignment and intefeence cancellation. Intefeence alignment aligns a subset of the ackets at the fist AP, allowing it to locally decode one acket and hence bootsta the decoding ocess. Intefeence cancellation enables othe APs to use the decoded acket to cancel its intefeence, and hence decode moe ackets. Neithe intefeence alignment no cancellation would be sufficient on its own to decode the thee ackets in Fig.. 9

2 IAC has the following featues: IAC bings in moe gains than aaent in the above examle and genealizes to any numbe of antennas. Fo a MIMO system with M antennas, we ove analytically that IAC delives M concuent ackets on the ulink, and max(m, M ) on the downlink i.e., it doubles the thoughut of the ulink, and almost doubles the thoughut of the downlink fo a lage numbe of antennas. IAC delegates all coodination to the APs, which tell the clients how to encode thei ackets to oduce the desiable alignment. Futhe, the channel estimates equied fo comuting this alignment can be comuted fom ack ackets with negligible ovehead. IAC woks with vaious modulations and FEC codes. This is because IAC subtacts intefeence befoe assing a signal to the est of the PHY, which can use a standad 80. MIMO modulato/demodulato and FEC codes. We have built a ototye of IAC in GNU-Radio and evaluated it using a testbed of 0 USRP nodes, each equied with -antennas. Ou esults eveal the following findings: IAC imoves the aveage thoughut of ou 0-node -antenna MIMO LAN by.x on the downlink and.08x on the ulink. These exeimental gains ae slightly highe than the analytical ones because ou analysis does not model IAC s divesity gains. IAC is fai in the sense that evey client in ou testbed benefits fom using IAC instead of cuent MIMO. IAC ovides a gain fo any numbe of clients including a single active client. In this case, IAC exloits divesity to imove the thoughut by.x.. Contibutions This ae makes thee main contibutions: It esents intefeence alignment and cancellation (IAC), a new intefeence management technique that synthesizes intefeence alignment and intefeence cancellation, showing that the combination inceases the thoughut in scenaios whee neithe alignment no cancellation alies seaately. It analytically demonstates that IAC almost doubles the multilexing gain (i.e., numbe of concuent tansmissions) of flat-fading intefeence-limited MIMO LANs. The caacity of a distibuted netwok can be witten as [6]: C(SNR) = dlog(snr) + o(log(snr)), whee d is the multilexing gain and the caacity is comuted as a function of the signal to noise atio (SNR). At elatively high SNRs, the caacity is dominated by the fist tem and linealy inceases with the multilexing gain, d. We ove that IAC inceases the multilexing gain of flat-fading MIMO LANs, and thus ovides a linea incease in the caacity chaacteization of these netwoks. It esents the fist imlementation of intefeence alignment demonstating its feasibility. Ou esults show that in flat-fading channels, alignment can be efomed without any synchonization even in the esence of diffeent fequency offsets between concuent tansmittes. Related Wok Related wok falls in the following aeas. (a) MIMO Communication Theoy. Ou wok builds on the theoy of intefeence alignment. Recent wok has agued that e-ocessing signals at the sendes in a manne that aligns intefeence at the eceives inceases the total caacity of wieless netwoks [, 6, 9, ]. Howeve, to the best of ou knowledge, this ae is the fist to esent a system design and an imlementation of intefeence alignment, showing that such idea woks in actice. Futhe, this ae is the fist to combine intefeence alignment with intefeence cancellation, showing that the combination, temed IAC, inceases the thoughut in scenaios whee neithe alignment no cancellation hels alone. Ou wok builds on ecent advances in the theoy of multiuse MIMO (MU-MIMO). MU-MIMO advocates having multile clients concuently communicate with a single AP o base station [,, 9, 0]. Thus, the thoughut of MU-MIMO is limited by the numbe of antennas on a single AP []. In contast, this ae shows that IAC ovecomes the antennas-e-ap thoughut limit. Ou wok is also elated to Vitual MIMO [9, 0]. Vitual MIMO allows multile tansmittes to tansmit concuently and makes the eceives collaboate to jointly decode the concuent tansmissions. Vitual MIMO, howeve, emains a theoetical concet with no actical design because of two difficulties. Fist, it equies the tansmittes to be synchonized to the symbol level. Second, it equies the eceives to communicate the aw eceived signal samles to be jointly decode. Communicating signal samles geneates excessive ovehead because to catue a signal without loss of infomation one needs to samle it at twice its bandwidth at each antenna, with each samle about 8-bit long. Fo examle, to jointly decode thee APs with fou antennas each, one needs to send 6 Gb/s on the Ethenet. In contast, IAC s eceives communicate decoded ackets, and hence the Ethenet taffic emains comaable to the wieless thoughut. (b) Wieless Netwoks. Past wok on single-antenna systems has oosed using multile APs to imove coveage [6, 8], balance the load [], o ecove couted ackets [, ]. This ae use multile APs but focuses on MIMO netwoks, and intoduces IAC, a new technique that enables MIMO LANs to suot a lage numbe of concuent tansmissions than ossible with existing designs. Pio wok has also advocated allowing concuent tansmissions in the context of single-antenna nodes. Some of these designs event intefeence by dividing the esouces between uses. Fo examle, they might assign the diffeent uses diffeent fequency bands [, 6], o diffeent codes [7, 7]. Othe designs use intefeence cancellation to decode in the esence of intefeing signals [, 8]. IAC diffes fom this wok in focus because it addesses MIMO netwoks. It also diffes in mechanisms because IAC does not assign uses diffeent fequency bands o diffeent codes and alies to scenaios whee intefeence cancellation alone does not aly. Finally, APs with diectional antennas divide the sace into sectos, each seved by a diffeent antenna. This events intefeence between nodes in diffeent sectos, allowing multile clients to communicate concuently with the AP. Ou aoach is othogonal to diectional antennas since we can enable nodes in the same secto (i.e., nodes that intefee) to communicate at the same time. Intefeence Alignment and Cancellation IAC s design tagets MIMO wieless LANs in a univesity o cooate camus whee APs ae connected via a wied infastuctue (e.g., Ethenet). Today these netwoks use one AP to seve any aticula aea, and limit intefeence by assigning adjacent APs to diffeent It is a common mistake to think that MIMO beam-foming is equivalent to diectional antennas. Beam-foming allows the signal to constuctively combine at the intended eceive, inceasing its thoughut. This howeve still ceates intefeence at nodes that ae not in the diection of the intended eceive. Hence, beam-foming cannot ovecome the antennas-e-node thoughut limit of MIMO LANs. 60

3 Client h h h h H = h h h h AP H 0 0 H Figue : Two Packets on Ulink. The client tansmits two ackets, and, fom its two antennas. The ackets aive along the vectos H[ 0] T and H[0 ] T, whee H is the channel matix and [.] T efes to the tansose of a vecto. To decode and, the AP ojects along the vectos othogonal to H[0 ] T and H[ 0] T esectively. 80. channels. Simila to the cuent achitectue, in IAC, adjacent aeas emloy diffeent 80. channels, but in contast to the cuent achitectue, each of these aeas is seved by a set of APs on the same channel, athe than a single AP. IAC allows this set of APs to seve multile clients at the same time desite intefeence. To do so, it leveages the wied bandwidth to enable the APs to collaboate on esolving intefeing tansmissions. IAC has thee comonents: ) a hysical laye that decodes concuent ackets acoss APs, ) a MAC otocol that coodinates the sendes to tansmit concuently on the wieless medium, and ) an efficient mechanism to estimate channel aametes. IAC s Physical Laye IAC modifies the hysical laye to allow multile client-ap ais to communicate concuently on an 80. channel. IAC oeates below existing modulation and coding and is tansaent to both. Fo claity, we esent ou ideas in the context of a -antenna enode system, and assume nodes know the channel estimates. Late, we extend these ideas to any numbe of antennas and exlain how we measue channel functions. Ou esentation focuses on scenaios whee intefeence fom concuent tansmissions is much stonge than noise and is the main facto affecting ecetion. (a) Two concuent ackets on the ulink: Let us stat with the standad MIMO examle in Fig., whee a single client tansmits two concuent ackets to an AP. Say that the client tansmits on the fist antenna, and on the second antenna. The channel linealy combines the two ackets (i.e., it linealy combines evey two digital samles of the ackets). Hence, the -antenna AP eceives the following signals: y = h + h y = h + h, whee h i j is a comlex numbe whose magnitude and angle efe to the attenuation and the delay along the ath fom the i th antenna on the client to the j th antenna on the AP, as shown in Fig.. Since the nodes have two antennas, the tansmitted and eceived signals live in a -dimensional sace. Thus, it is convenient to use - dimensional vectos to eesent the system [9]. This eesentation will allow us to use simle figues to descibe how a MIMO system woks. We can e-wite the above equations as: ( ) y y = H ( 0 ) + H ( 0 ), () whee H is the ulink channel matix (i.e., the matix of h i j s). Thus, the AP eceives the sum of two vectos which ae along the diections H[ 0] T and H[0 ] T (whee [.] T efes to the tansose of a vecto), as shown in Fig.. + H H H 0 H 0 H 0 H H 0 0 H H H Clients APs 0 (a) Thee Packets Without IAC. v v v H H H H H v Hv H v Hv Clients APs H v (b) Thee Packets With IAC. H Figue : Thee Packets with/without IAC. In (a), the clients tansmit the ackets without alignment. The ackets combine at the APs along thee diffeent vectos and the APs cannot decode any acket. The second case shows how IAC delives thee ackets on the ulink. Secifically, two of the thee ackets ae aligned at AP, allowing AP to decode one acket and send it to AP on the Ethenet. AP uses intefeence cancellation to subtact the acket and decode the emaining two ackets. Assume the AP knows the channel matix, H, (we will see how to estimate it in 8). Decoding is easy; to decode, the AP needs to get id of the intefeence fom, by ojecting on a vecto othogonal to H[0 ] T. To decode it ojects on a vecto othogonal to H[ 0] T. We efe to the diection that a eceive ojects on, to decode, as the decoding vecto. (b) Thee concuent ackets on the ulink: Conside what haens if anothe client concuently tansmits a acket, as shown in Fig. a. Using the same deivation as above, AP eceives: ( ) ( ) ( ) ( ) y 0 = H y 0 + H + H 0, whee H and H ae channel matices fom the fist and second clients to AP. Said diffeently, AP eceives the combination of thee ackets,, and, along thee vectos H [ 0] T, H [0 ] T and H [ 0] T, as shown in Fig. a. Since AP has only two antennas, the eceived signal lives in a -dimensional sace; hence AP cannot decode thee ackets. Said diffeently, fo any acket i, the AP cannot find a ojection (decoding vecto) that eliminates intefeence caused by the othe two ackets. The second access oint, AP, is in a simila state, it eceives thee ackets along thee vectos H [ 0] T, H [0 ] T and H [ 0] T, and cannot decode fo the same eason. Howeve, one advantage of MIMO is that a tansmitte can contol the vectos along which its signal is eceived. Fo examle, when a tansmitte tansmits acket on the fist antenna, this is equivalent to multilying the samles in the acket by the unit vecto [ 0] T befoe tansmission. As a esult the eceived vecto at the AP is H[ 0] T, whee H is the channel matix fom tansmitte to eceive. If the tansmitte, instead, multilies the acket by a diffeent vecto, e.g., v, the AP will eceive the vecto H v. Thus, instead of tansmitting each acket on a single antenna, we multily acket i by a vecto v i (i.e., multily all digital samles in the acket by the vecto) and tansmit the two elements of the esulting -dimensional v 6

4 + v v Hv Hv v H H H v v v v Clients APs H v H v H v H v Hv H v Hv Figue : Fou Packets on the Ulink. IAC allows AP and AP to decode one acket each, and AP to decode the two emaining ackets. This equies thee ackets to be aligned at AP and two ackets at AP, which can be done by icking aoiate encoding vectos. vecto, one on each antenna. Thus, by changing v i, we can contol the vecto along which the AP eceives the acket. We call the vecto v i the encoding vecto of acket i. Now, we can aly this method to the -client and -AP system to tansmit thee concuent ackets. In aticula, the tansmittes multily acket i with vecto v i, as shown in Fig. b. We want to ick v and v such that the second and thid ackets (i.e., and ) ae aligned at AP, as in Fig. b, that is: H v = H v, () whee H and H ae the channel matices fom the fist and second clients to AP. This can be easily done by icking andom (but unequal) values fo v and v and substituting in the above equation to get v (i.e., v = H H v ). In this case, AP eceives the second and thid ackets aligned on the same diection as in Fig b. Thus, AP can decode the fist acket,, by ojecting on a vecto othogonal to the aligned intefeence, i.e., a vecto othogonal to H v and H v. Since these two vectos ae aleady aligned, thee is a vecto that is othogonal to both of them, and thus the AP can decode. Note that without alignment, AP could not decode because H v and H v would have diffeent diections, and no vecto will be othogonal to both. Note that aligning two vectos with esect to AP does not mean that they ae aligned with esect to AP. This is because the channels fom the clients to the two APs ae diffeent and indeendent. Howeve, we do not need to align the signals at AP. AP can decode the fist acket and send it to AP on the Ethenet. Now AP knows the fist acket. It also knows the channel functions (see in 8 how we comute channel functions). Hence it can econstuct the signal associated with the fist acket and subtact it fom what it eceived. This is standad intefeence cancellation [9, 9]. Afte cancellation, AP is back into a scenaio simila to tyical MIMO, namely two ackets on two diffeent diections, in a -dimensional sace. Hence, it can decode. Thus, we obtained all thee ackets. AP decoded the fist acket, and AP decoded the second and thid. (c) Fou concuent ackets on the ulink: Let us ty to incease the numbe of concuent ackets on the ulink to. We cannot do this with only clients and APs (This is because the system In geneal, aligning the diections would mean H v = αh v, whee α is a scala. Also note that the vectos ae nomalized to satisfy the owe constaints. But fo claity, we ignoe these details in ou descition. Channel matices ae tyically invetible because the antennas ae chosen to be moe than half a wavelength aat. If the matix is not invetible, then you don t eally have a MIMO system because the two antennas tanslate into just one equation. is aleady too constained to oduce the desiable alignment.) We need to add an additional AP-client ai. Fo examle, conside the thee APs and thee clients, in Fig.. The fist client tansmits ackets and, the second client tansmits and the thid client tansmits the fouth acket,. Now that we have develoed a vecto eesentation, it is faily simle to oduce an IAC solution fo any configuation. Secifically, as shown in Fig., AP needs to align out of ackets. This esults in one fee acket, e.g.,, which can be decoded with othogonal ojection, as we did ealie. Fom the esective of AP, is aleady decoded at AP, and hence can be subtacted and emoved fom the signal. Thus, AP is left with thee unknown ackets. To decode one moe acket, it needs to have out of ackets aligned, as shown in Fig.. Fom the esective of AP, two ackets ae aleady decoded at AP and AP, and thei signal can be canceled using intefeence cancellation. Thus, AP is left with only two unknown ackets, which it can decode. Hence, AP does not need to align any ackets. We can achieve the desied alignment (i.e., the alignment in Fig. ) by solving the following equations: H v = H v = H v () H v = H v, () whee H i j is the ulink channel matix fom the i th client to the j th AP. Eqs. ensues the desied alignment at AP and Eq. ensues the desied alignment at AP. Effectively, this tanslates to thee linea equations in thee unknowns (the vectos), which can be solved. Thus, the APs can decode fou concuent ackets. (d) The downlink: The discussion so fa has focused on the ulink, what about the downlink? Clealy the downlink is moe limited, since the clients cannot cooeate ove a wied Ethenet. A client cannot decode one acket and send it to othe clients fo intefeence cancellation. The lack of cooeation means that the clients have to decode indeendently. So, we need to align the intefeence at each client to ensue that it can decode at least one acket. Fo a -antenna system, this means that we can at best delive concuent ackets on the downlink. This howeve is still highe than what can be deliveed in today s oint-to-oint MIMO LANs. Say that we want to delive ackets,, and to Client, Client, and Client esectively. Each client needs to eceive the two undesied ackets aligned along the same vecto and the desied acket along a diffeent vecto, as shown in Fig. 6. To achieve this behavio, each AP tansmits one of the thee ackets. Now the oles ae flied: the APs ae the tansmittes and the clients the eceives. Hence, each AP multilies the tansmitted acket by a vecto v i that is caefully chosen to ensue the desied alignment. Secifically, we need to ensue: H d v = H d v () H d v = H d v (6) H d v = H d v, (7) whee H d i j is the channel fom the ith AP to the j th client, i.e., the downlink channels. The thee equations above align the ackets at each client to ensue that the two undesied ackets ae along the same vecto. These ae thee linea equations ove thee unknown vecto and can be solved using standad methods (simila to how we solved Eqs. and ). Hence, each client can decode its desied acket by othogonal ojection. The solution to the alignment is v = eig(h H H H ), whee eig(h) is an eigen vecto of H, and v = H H v and v = H H v. 6

5 Hv d Hv d Hv d H v d H v d Hv d H v d Hv d H v d v v v Clients APs Figue 6: Thee Packets on the Downlink. The APs delive one acket to each client. To enable the client to decode its acket, all the undesied ackets at the client must be aligned. Clients Figue 7: Fou Packets on Downlink. At the fist client, ackets and ae aligned along one dimension, allowing and to lie in a two dimensional sace and hence be decoded. Similaly, at the second client, ackets and ae aligned, allowing and to be decoded. Beyond Two Antennas The evious section focuses on -antenna systems, but fo the geneal case of M antennas e-node, what is the maximum numbe of concuent ackets that can be deliveed? Futhe, how many APs ae needed to suot such a system? Naively, it might seem that the numbe of concuent ackets is constained only by the numbe of APs. Secifically, it might seem that one can align the eceived ackets at evey AP, allowing each of them to decode at least one acket, and hence one can kee inceasing the numbe of concuent ackets by inceasing the numbe of APs. This is howeve misleading because aligning a signal at one eceive limits the ability of the tansmitte to feely align it at a second eceive. In aticula, evey alignment imoses new constaints on the encoding vectos at the tansmitte. Fo a feasible solution, the constaints should stay fewe than the fee vaiables in an encoding vecto. Since the encoding vecto has as many vaiables as thee ae antennas on the node, the numbe of constaints cannot exceed the numbe of antennas. Thus, using moe APs is beneficial but only u to a oint, afte which one needs to incease the numbe of antennas. Below, we demonstate that in IAC, the numbe of concuent ackets can be almost twice the numbe of antennas, and that this gain is achieved with a elatively small numbe of APs. (a) Downlink. In [], we ove the following: Lemma. In a system with M antennas e node, the maximum numbe of concuent ackets IAC can delive on the downlink is max{m, M }. Fo M >, IAC achieves this with M APs. Fo M =, the above lemma tells us that we can achieve concuent ackets on the downlink. Fig. 7 shows the downlink case. We have two APs and two clients. Each AP tansmits two ackets, one fo each client. Since the clients have thee antennas, the signal is in a AP 6 Clients AP 6 6 Figue 8: Six Packets on Ulink. At AP, all the ackets othe than ae aligned on a two dimensional lane, allowing to be decoded. At AP,, and 6 ae aligned along, allowing and to be decoded. At AP, we cancel, and leaving, and 6 to lie along thee diffeent dimensions and be decoded. thee dimensional sace. Thus, if we align two ackets along one dimension, the othe two ackets ae fee of intefeence and can be decoded, esulting in concuent ackets. The above ocedue can be genealized to any numbe of antennas. Secifically, if we have M APs and two clients, a ocedue that makes each AP tansmit a acket to each client can delive a total of M concuent ackets acoss the two clients. Fo a lage M, this almost doubles the thoughut of cuent MIMO LANs. (b) Ulink. In [], we ove the following: Lemma. Fo a M-antenna system, thee o moe APs, and at least two clients, IAC can delive M concuent ackets on the ulink. Fo M =, the above lemma tells us that we can achieve 6 concuent ackets on the ulink. Fig. 8 shows thee clients tansmitting to thee APs. At the fist AP, five out of six ackets ae aligned in the same lane. This leaves one acket fee of intefeence and hence can be decoded. Fom the esective of the second AP, one acket is aleady decoded and hence can be eliminated fom the eceived signal. Out of the five ackets left, the second AP needs to have thee ackets aligned along one dimension and two fee ackets, allowing it to decode two ackets. Finally, fom the esective of the last AP, thee ackets ae aleady decoded and hence thei intefeence can be eliminated. This leaves the last AP with thee unknown ackets in a thee dimensional system and hence it can decode all of them. Again, this ocedue can be alied indeendent of the numbe of antennas. Secifically, one needs to align M ackets such that the fist AP can decode one acket, the second AP decodes M ackets and the last AP decodes M ackets. 6 Pactical Issues The acticality of IAC elies on being able to imlement intefeence alignment and intefeence cancellation. IAC uses only the subtaction ste of intefeence cancellation. Intefeence cancellation tyically involves two stes: fist it decodes one of the concuent ackets in the esence of intefeence and second it subtacts the decoded acket fom the est to emove its contibution to intefeence, allowing the decoding of moe ackets. IAC elaces the fist ste with intefeence alignment to othogonalize intefeence and eliminate its imact as it decodes one of the concuent ackets. It uses intefeence cancellation only to subtact the decoded acket. The subtaction ste of intefeence cancellation is widely studied and has been shown to wok in actical imlementations [,, 8]. 6 6

6 Futhemoe, the subtaction ste does not equie any synchonization between tansmittes, woks with OFDM systems and vaious modulation schemes, and can accommodate single ta and multi-ta fequency selective channels [9, 0]. In contast, io to this ae, intefeence alignment has been a uely theoetical idea with no actical imlementation. Thus, in this section, we focus on the acticality of efoming alignment. (a) Fequency offset: In actice, a tansmitte-eceive ai always exhibits a small fequency offset, f. The fequency offset causes the hase of the eceived signal to incease linealy with time, i.e., the eceived vecto otates with time. Since the fequency offset is tyically diffeent fo diffeent sende-eceive ais, signals fom diffeent tansmittes that ae aligned at the same eceive will otate at diffeent ates. Thus, it might seem that signals that ae aligned at the beginning of a acket will lose alignment with time and be comletely misaligned by the end of the acket. This easoning howeve is incoect because intefeence alignment haens in the antennasatial domain and not the I-Q domain. 6 Diffeences in fequency offset cause elative diffeences in how the signals otate in the I-Q domain but only scale the diection of the vectos in the satial domain by a comlex numbe, leaving the alignment unaffected. Secifically, suose the encoding vectos, v and v, ae icked to satisfy the equation H v = H v. As a esult of the two fequency offsets, f and f, the channel, H i (t), changes as a function of time as H i e jπ f it. Thus, these time vaying channels satisfy the equation: H (t)e jπ f t v = H (t)e jπ f t v H (t) v = e jπ( f f )t H (t) v The comlex function e jπ( f f )t scales the vecto, H (t) v, leaving its oientation unaffected. Since alignment only equies that the two vectos have the same oientation, the signals emain aligned though the end of the ackets desite diffeent fequency offsets. Realizing that signal alignment is unaffected by otation in the I-Q domain is an imotant lesson that we leaned fom the imlementation. (b) Diffeent Modulations: Intefeence alignment woks indeendent of what constitutes the signal, i.e., indeendent of the modulation scheme (BPSK, QAM, o OFDM). It might seem that the modulation scheme, say QAM, changes the signal oientation and hence beaks the alignment. Again this agument is incoect because modulation changes the signal s oientation in the I-Q domain, but intefeence alignment haens in the antenna satial domain. (c) Symbol Synchonization: One lesson that we leaned fom the imlementation is that fo elatively flat channels, you do not need to have symbol level synchonization. Secifically, if the channel between each tansmit-eceive antenna ai can be eesented by a single comlex numbe, h i j, whose magnitude efes to the attenuation and hase efes to the delay along the ath, intefeence alignment can then be imlemented accuately without tansmitte synchonization. This aises fom two facts: ) we efom intefeence alignment at the signal level and not symbol level, i.e., we align signal samles egadless of what symbol they eesent, ) the alignment occus in the satial antenna domain, not the I-Q domain, and hence though unsynchonized tansmittes may not be aligned in the I-Q domain, this does not affect thei alignment in the satial antenna domain. 7 Once the eceive knows the bits and estimates the channel function fom the eamble, it can econstuct the coesonding continuous signal, samle it at the desied oints, and subtact it fom its eceived vesion. 6 The I-Q domain is the -dimensional sace that efes to the tansmitted comlex numbe. 7 It should be noted that intefeence alignment is diffeent fom multi-use MIMO (which tyically equies synchonization) in that not all signals need be decodable at a eceive. Secifically aligned intefees need not be decodable. Note that modeling the channel between a ai of antennas as a single comlex numbe is accuate fo naowband o flat channels, but becomes less so as the width of the channel inceases. We conjectue that even if the channel is not quite flat, one can still do the alignment seaately in each OFDM subcaie without tying to synchonize the tansmittes. In this case, thee is some intefeence between the OFDM subcaies, but given that neaby subcaies tyically have simila fequency esonse, fo modeate width channels the esulting imefection in the alignment stays accetable. We cannot check this conjectue on USRP since thei channel is faily naow and is accuately modeled with a single comlex numbe. 7 Medium Access Contol Since IAC allows multile clients and APs to tansmit simultaneously, it changes the equiements of the MAC. The challenge in designing a MAC otocol fo IAC aises not only fom the need to enable multile nodes to concuently access the medium, but also fom ou desie to maintain minimal comlexity at the clients. Secifically, a client should be oblivious to the numbe of APs in the system, and othe clients who tansmit concuently. Finally, since taffic is busty, we need to dynamically change the combination of concuent clients to match instantaneous taffic demands, while esecting fainess. The basic incile undelying ou solution is to move comlexity to APs, which abitate the medium among clients, and also ovide each client with its encoding and decoding vectos. Ou solution has two comonents: ) a MAC otocol that allows multile nodes to access the medium concuently, and ) a concuency algoithm that decides which clients uload/download concuently. 7. Accessing the Medium Ou design extends the 80. Point of Coodination Function (PCF) mode to allow it to suot multile concuent sendes. PCF is at of the standad []. It allows the AP to abitate the medium by olling the clients, and is oiginally designed to enable 80. netwoks to deal with time sensitive infomation. (a) Contention-Fee and Contention Peiods. In IAC, one of the APs is designated as the leade. The leade AP acts as a coodinato. It olls the clients and gants access to those who have data to tansmit []. Simila to PCF, we divide time into: Contention Fee Peiod (CFP) and Contention Peiod (CP), as shown in Fig. 9. A contention-fee eiod stats with the leade AP boadcasting a beacon that announces the duation of the cuent CFP. Duing a CFP, the leade AP coodinates access to the medium enabling the nodes to tansmit using IAC. This is followed by a contention eiod, duing which any node can contend fo the channel using standad 80.n. The objective of this design is to use the contention eiod to allow new clients to associate with the APs, o to tansmit afte a long eiod of silence, using oint-to-oint MIMO. In contast, the contentionfee eiod (CFP) is used to ack tansmissions as much as ossible, inceasing thoughut. The duation of the contention eiod (CP) is constant, while the duation of CFP vaies deending on congestion. Duing CFP, the APs seve one acket (on ulink and downlink) to each client that has ending taffic. Hence, when congestion is low and queues ae emty, the CFP natually shinks, and clients send moe time in CP. When congestion is high, many clients have ending taffic and hence the CFP exands, which is desiable as this mode uses IAC to ack tansmissions and incease efficiency. (b) Acquiing Medium Duing CFP. Next, we exlain how concuent tansmittes acquie the medium duing a CFP. Clealy, this 6

7 Data+Poll Data+Poll Data+Poll Beacon + ACK Ma ACKs Data+Poll Data+Poll Data+Poll Downlink Fid #APs Contention-fee eiod.... Data+Req Data+Req Data+Req Ulink Data+Req Data+Req Data+Req Figue 9: IAC s Extension to PCF CF-End CRC Contention Peiod CF- End Figue 0: The metadata in a DATA+Poll fame. This metadata is boadcast by the leade AP alone to infom the clients in a downlink tansmission gou of thei decoding vectos and the othe APs of thei encoding vectos. equies knowing which clients ae seved concuently. This is the job of the concuency algoithm, which divides clients with ending taffic into gous of concuent tansmissions that we call tansmission gous. It futhe decides which AP seves which client in a tansmission gou, and the values of the encoding and decoding vectos. The ocess fo deciding this is descibed in 7.. In this section, we focus on how to delive ackets in each tansmission gou. Fig. 9 shows the seies of events duing a contention-fee eiod. At the beginning of a CFP, the leade AP sends a beacon. The leade AP then stes though the downlink tansmission gous, one at a time, tansmits thei downlink ackets with the hel of othe APs, and olls the coesonding clients fo ulink taffic. This mode is simila to cuent PCF behavio, excet that in the cuent PCF the AP stes though a list of individual clients, one at a time; wheeas in IAC the leade AP stes though a list of tansmission gous. (b.) Downlink. The leade AP fist goes though the list of downlink tansmission gous. With the hel of othe APs, it sends a DATA+Poll fame to each gou. This fame has two ats. The fist at, shown in Fig. 0, is boadcast by the leade AP alone, and contains the ids of the clients in the gou and thei encoding and decoding vectos. The ids ae given to the clients uon association. The encoding and decoding vectos ae comuted by the concuency algoithm which uns on the leade AP. The leade AP also includes a fame id,fid, the numbe of APs and a checksum of its boadcast. Futhe, it sets the length of thedata+poll fame to the maximum length of the ackets in the tansmission gou, so that all clients know when the fame ends. The second at of the fame is the combination of concuent tansmissions by all APs. Fo the examle of thee APs with -antennas each, this at has the thee APs tansmitting a acket to each of the thee clients in a tansmission gou. Note that both the clients and APs listen to the leade AP as it boadcasts the fist at of the DATA+Poll fame. In ode to tansmit concuently, the APs need to lean thei encoding vectos. Similaly, the clients need to lean the decoding vectos to be able to decode thei data. The clients and APs can use the checksum to test whethe they eceived the coect infomation. Note that the tansmissions still wok fine if any of the APs o the clients failed to hea the leade AP. Secifically, the AP/client who failed to hea the leade AP, will not tansmit. The othe tansmissions can go as desied. Afte the DATA+Poll fame, the clients in the tansmission gou send thei acks, one afte the othe, using taditional MIMO. The ode in which they tansmit these acks is the same as the ode of thei ids in the DATA+Poll fame. These acks ae simila to synchonous 80. acks. In 80., they ae sent one afte each data acket. Hee the data ackets ae sent concuently and all acks follow. (b.) Ulink. Afte going though all downlink gous, the leade AP stes though the ulink gous. Simila to the downlink case, the leade AP fist boadcasts agant fame secifying the ids of the clients that will tansmit on the ulink, and the encoding and decoding vectos. The othe APs listen to the encoding and decoding vectos and wait fo clients tansmissions. The clients in an ulink gou use thei encoding vectos to tansmit simultaneously on the ulink. Each client tansmits a Data+Req fame. This fame contains the client s ulink data. If the client still has taffic to send, the fame will also contain a new equest fo tansmission. Each AP listens to thedata+req fame and ojects the eceived signal on the oe decoding vecto. This ojection is othogonal to the intefeing signals and hence it allows each AP to eceive its client of inteest. One diffeence between the ulink and downlink is that, while each client on the downlink can immediately ack its acket, the APs need to decode successively using intefeence cancellation and hence cannot send synchonous acks. The solution howeve is simle. Duing the following contention eiod, the APs infom the leade AP of successful ecetions using Ethenet. The leade AP combines and sends all acks at the beginning of the next CFP, by embedding them in the beacon infomation as a bit ma. This should not cause any significant delay since it allows all clients in the CFP mode to lean about thei evious acket befoe they get to send the next acket. At the end of CFP, the leade AP sends a CF-End fame. This allows the clients to go back to the contention mode, whee they use taditional oint-to-oint MIMO. A few oints ae woth noting. (a) How do we deal with lost ackets and etansmissions? If a acket is lost on the ulink, the client discoves the loss fom the lack of an ack (at the beginning of the next CFP) and asks fo a new tansmission slot next time it is olled. On the downlink, the coesonding AP discoves the acket loss immediately, fom the lack of a client ack, and asks the leade AP to schedule a etansmission. (b) Is it ossible fo vaious APs to make inconsistent decisions? Only the leade AP makes decisions, while othe APs ae dumb tansmittes/eceives. Simila to clients, they eceive thei encoding and decoding vectos fo each tansmission gou ove the medium and use them without any modification. They only infom the leade AP in case a acket is lost, o the channel s estimate has changed. (c) How often do APs need to communicate ove the Ethenet and what do they exchange? As descibed in, APs exchange the decoded ackets ove the Ethenet to efom intefeence cancellation. Futhe, the subodinate APs need to tell the leade AP wheneve a acket is lost o channel coefficients to a client changes by moe than a theshold value. The APs can send this infomation as annotation on ackets they exchange to efom cancellation. (d) How lage is the Ethenet ovehead? To minimize Ethenet ovehead, IAC connects the set of APs using a hub. This design ensues that evey decoded acket is boadcast only once to all APs and to the switch that fowads the acket to its wied/final destination. In this design evey acket is tansmitted once and thee is no exta ovehead. While a hub is less efficient fo a geneal Ethenet than a switch, it is a natual choice to connect the IAC APs. This hubbed netwok is then connected to the est of the Ethenet via a switch. (e) How lage is the wieless ovehead associated with IAC s MAC? IAC intoduces metadata to coodinate clients and APs. 6

8 Secifically, concuent tansmissions ae eceded by a shot boadcast fom the leade AP to infom the client-ap ais of thei encoding and decoding vectos. Such a boadcast message aleady exists in 80. PCF mode. 8 We only annotate these messages with exta infomation that is a few bytes e client-ap ai. Assuming 0 byte ackets, the ovehead of the metadata amounts to -%. In comaison, the thoughut imovement exected fom IAC is.x to x, which moe than comensates fo the loss. 7. Concuency Algoithm The concuency algoithm uns at the leade AP. The leade AP maintains a FIFO queue fo taffic ending fo the downlink and a simila queue fo ulink equests leaned fomdata+poll fames (see 7.). Given the queues of ulink and downlink taffic, the concuency algoithm geneates the ulink and downlink tansmission gous. Without loss of geneality, we will focus on the downlink. Thee ae multile otions fo how to combine clients. The bute foce aoach consides all combinations of clients with queued ackets and all diffeent ways of assigning them to existing APs, comutes the encoding and decoding vectos, and estimates the thoughut of each combination. The thoughut of a tansmission gou can be estimated without any tansmissions as: i log(+ v T i H i w i ), whee the sum is ove client-ap ais, H i is the channel fo a ai, and v i and w i ae the coesonding encoding and decoding vectos [9]. It then ceates tansmission gous fo the queued ackets which maximize thoughut. Thee ae two oblems with such an aoach. Fist, estimating the thoughut fo evey combination of clients in the queue is a combinatoial oblem in the numbe of clients. Second, since this aoach focuses on maximizing thoughut, it always efes clients with good channels and hence is unfai. Altenatively, one can always ceate tansmission gous by combining ackets accoding to thei aivals in the FIFO queue. This aoach is simle and gives each client a fai access to the medium, but is oblivious to the thoughut of a aticula gouing. In actice, diffeent gous may yield significantly diffeent thoughut gains (see 0.). (a) The Best of Two Choices. IAC s concuency algoithm balances the desie fo high thoughut with the need to be fai. To event stavation and educe delay, it always icks the head of the FIFO queue as the fist acket in the cuent tansmission gou. To educe comutational ovehead, it icks othe clients in the gou using the best of two choices, a standad aoach fo educing the comlexity of combinatoial oblems []. Say each gou has thee clients, and we aleady icked the fist client in the gou as the client whose acket is at the head of the tansmission queue. We andomly ick two clients with queued ackets as candidates fo the second osition in the gou. Similaly, we also andomly ick two clients fo the thid osition in the gou. Now we estimate the thoughut fo the fou tansmission gous fomed by these otential candidate clients and ick the gou that otimizes thoughut. As a esult, instead of comuting thoughut fo evey ossible combination of clients, we just comute it fo fou andom client combinations. Let us now conside the fainess of the aoach. A client is consideed fo tansmission eithe because it is at the head of the queue o because of a andom choice. Both these cases give the client a fai access to the medium. Howeve, since afte icking the candidate clients, we still otimize fo thoughut, we need a mechanism to ensue that clients that neve maximize thoughut get icked. To do this, we assign a cedit counte to each client. If the client is consideed as a esult of a andom choice, and is ignoed since calls thegant famecf-poll, i.e., it is a oll without downlink data. it does not maximize thoughut, the counte is incemented; but if it is icked fo tansmission the counte is eset. If the counte cosses a theshold, the client is selected as at of the gou iesective of the thoughut. This mechanism ensues that evey client is at of some gou at least a minimum numbe of times. 8 Channel Estimation In IAC, the APs estimate and convey the channels to the leade AP as annotation on the decoded ackets sent ove the Ethenet. (a) Ulink: To estimate the encoding and decoding vectos fo the ulink, we need the hysical channel fom each concuent client to each AP, as shown in Eqs. and. In the absence of concuent clients, estimating this channel is a standad MIMO technique []. Thus, the fist time a client boadcasts an association message, all APs estimate the channel fom that client to themselves. Once the APs have an initial estimate, they need to tack it. This is done using the client s ack ackets fom the contention-fee eiod, and its data ackets fom the contention eiod. Both acket tyes ae tansmitted without any concuent tansmissions. Hence they can be ocessed using standad MIMO channel estimation []. Since the APs can estimate the channel fom evey ack the client tansmits, they obtain a fequent estimate of the channel. In static envionments the channel is elatively stable and can be easily tacked at this estimation fequency. Slight inaccuacy in estimating the channel only means that the intefeence is not fully eliminated afte alying the encoding and decoding vectos. As long as most intefeence is eliminated, the loss in thoughut stays negligible. (b) Downlink: Channel estimation is tyically done at the eceive [7]. Thus, we have two otions: eithe have clients estimate and convey the channels to the leade AP when it olls them, o ty to have APs estimate the channel by exloiting eciocity between ulink and downlink channels. In ou measuements, the latte otion woked with sufficient accuacy and hence we adot it. Reciocity means that the channel fom node A to node B is the tansose of the channel fom B to A. Thus, an AP can use the ulink channel fom a aticula client to infe the downlink channel to that client. It is imotant to undestand that channel eciocity does not mean that the link between two nodes A and B is symmetic. Reciocity (i.e., the kind that we cae about in this ae) means that the channel coefficients ae the same, but the noise o intefeence could be vastly diffeent. Fo examle, if A tansmits symbol x, node B eceives y B = Hx+n B. Similaly, in the oosite diection, node A eceives y A = Hx+n A. The channel multilie, H, is the same, but the noise could be much highe at A if it is close to a micowave oven. Hence, one may see many acket dos at A but not at B, but this does not contadict eciocity. Reciocity has been confimed in measuements [6, 8, 7] and is used in QUALCOMM s 80.n oosal []. Reciocity cannot be alied diectly without calibation to account fo hadwae diffeences between the tx and x chains. The calibation howeve can be comuted once and does not change fo the same sende eceive ai. IAC uses a calibation method fom the QUALCOMM s 80.n oosal []. Let H d be the channel between a aticula AP and client ai, and H u the ulink channel fom that client to the same AP. Then: (H d ) T = C Client,x H u C AP,tx, (8) whee H T efes to the tansose of H, and C Client,x and C AP,tx ae constant diagonal matices that descibe the exta attenuation and delay obseved by the signal in the tansmit and eceive hadwae chains on the client and the AP esectively. 66

9 Figue : Testbed Toology. 9 Comlexity IAC multilies each acket with an encoding vecto at the tansmitte and ojects on a diection othogonal to intefeence at the eceive. Both e-coding and ojection ae geneal oeations in MIMO designs []. IAC also efoms intefeence cancellation, which is linea in the numbe of cancelled ackets. Since, the ackets cancelled at an AP ae aleady decoded at io APs, all the ackets can be cancelled in aallel. Hence, the delay fom cancellation can be made indeendent of the numbe of cancelled ackets. 0 Pefomance Evaluation We evaluate IAC in a testbed of MIMO softwae adios. Each node is a lato connected to a -antenna USRP adio boad and uns the GNU-Radio softwae. To ceate a MIMO node, we equi each USRP with two RFX00 daughteboads. We also set the MUX value in softwae to allow the FPGA to ocess samles fom both antennas. (a) Toology. Ou testbed, shown in Fig., has 0 nodes. Each node has two antennas. All nodes ae within adio ange of each othe to ensue that concuent tansmissions ae enabled by the existence of multile antennas, not by satial euse. (b) Modulation. IAC uses the modulation/demodulation module as a black-box and hence woks with a vaiety of modulation schemes. Ou imlementation, howeve, uses BPSK, which is the modulation scheme that 80. uses at low ates. (c) Paametes. We use the default GNU-Radio aametes. Howeve, in ode to dive two antennas at the same time, we double the inteolation and decimation ates at the tansmitte and the eceive. Each acket consists of a -bit eamble, and 00-byte ayload. (d) Comaed Schemes. We comae the following: IAC: This is ou imlementation of IAC. 80.-MIMO: Thee ae multile oosals fo 80.n [, ]. These schemes ae all oint-to-oint, i.e., they allow only one tansmitte to access the medium at any oint in time. They howeve diffe in the amount of channel infomation available to the tansmitte, with moe channel infomation leading to bette efomance [9]. Since IAC uses full channel infomation, we comae it with an 80. MIMO design with full channel infomation available to both sende and eceive. This design is based on QUALCOMM s eigenmode enfocing [] and uses an aoach that is oven otimal fo oint-to-oint MIMO [9]. (e) Setu. In each exeiment, we andomly ick some nodes to act as APs and othes to act as clients. We eeat the same exeiment with IAC and 80.-MIMO. Thee oints ae woth noting. Fist, we allow 80.-MIMO access to the same numbe of APs as IAC. Though 80.-MIMO cannot use the additional APs fo concuent tansmissions, it can use them to incease divesity. Fo examle, if thee ae thee APs, each 80.-MIMO client communicates with the AP to which it has the best SNR. Second, we use a simlified TDMA MAC fo both IAC and 80.- MIMO. The MAC assigns the same numbe of tansmission timeslots to the two schemes. Conside an ulink scenaio that involves thee clients and thee APs. We stat with the 80.-MIMO exeiment and assign each client to its best AP. Each client tansmits fo 00 time slots, fo a total of 00 time slots fo the 80.-MIMO exeiment. We follow with an IAC exeiment whee clients tansmit togethe fo a total of 00 time slots. We then eeat the exeiment fo a diffeent client set. This simlified MAC allows fo a fai comaison between IAC and 80.-MIMO because it assigns the medium equally to each scheme. Imlementing the MAC in 7 equies access to accuate timing infomation, and the ability to quickly switch the boad fom a tansmit mode to a eceive mode. These equiements ae not suoted by the cuent USRP-GNU-Radio latfom. Finally, both IAC and 80.-MIMO use the GNU-Radio basic decoding modules (e.g., acket detection, clock ecovey, synchonization, and channel estimation) and the same system aametes. (f) Metic. It is tyical in the netwoking community to comae the thoughut of vaious designs. Thoughut esults, howeve, do not bing much insight fo adios that do not have oe ate adatation. Secifically, both in theoy and actice, wieless systems (e.g., 80.a/b/g/n cads, WiMax, etc.) can exloit a highe SNR to use dense modulation and coding schemes, and hence incease thei thoughut. GNU-Radios howeve do not yet suot ate adatation. In this case, it is not sufficient to comae thoughut because two systems may have the same thoughut yet one of them has a highe SNR. In an actual wieless oduct, the highe SNR system would use bette modulation and coding schemes to achieve a highe thoughut but cuent GNU Radios cannot exloit this highe SNR. Anothe way to look at the oblem is as follows. Say we take a -antenna system and show that IAC can decode fou concuent ackets, while 80.-MIMO decodes only concuent ackets. In this case, the thoughut of ou system will be double the thoughut of 80.- MIMO. Such a esult howeve is ambiguous because it is not clea whethe the 80.-MIMO system has a highe SNR. If it does, then 80.-MIMO could have used dense modulation and coding schemes, otentially doubling its thoughut, o maybe tiling it. Because of this ambiguity, it is efeable to measue efomance at the hysical laye in tems of SNR o a function of it. Thus, fo both 80.-MIMO and IAC, we measue the signal to noise atio, SNR Measued, fo each tansmitted acket. We comute the achievable ate, i.e., the ate that could be achieved in the esence of otimal ate adatation [9]: Rate = log (+SNR i Measued )[bit/s/hz], (9) i whee the sum is ove all concuent ackets. Fo each scheme, we aveage the above ate ove the whole exeiment, and comute the gain as the atio of the aveage ate of IAC to that of 80.-MIMO: Gain = 0. IAC s Multilexing Gain Rate IAC Rate 80. MIMO. (0) The main advantage of IAC is that it inceases the numbe of concuent ackets, i.e., it ovides a multilexing gain. In, we have demonstated this gain analytically. Hee, we check it in actice. Exeiment with -by- Ulink. We andomly ick two clients fom the testbed to uload taffic to two APs, then eeat the exeiment with diffeent clients and APs. We comae IAC to 80.- MIMO. In 80.-MIMO, each client uses its best AP and tansmits two ackets simultaneously, and the two clients altenate in using the medium. In IAC, the two clients simultaneously tansmit thee 67