A High Efficiency MAC Protocol for WLANs: Providing Fairness in Dense Scenarios

Transcription

1 A High Efficiency MAC Proocol for WLANs: Providing Fairness in Dense Scenarios Luis Sanabria-Russo, Jaume Barcelo, Boris Bellala, Francesco Gringoli arxiv:42.395v2 [cs.ni] Nov 5 Absrac Collisions are a main cause of hroughpu degradaion in WLANs. The curren conenion mechanism used in IEEE 82. neworks is called Carrier Sense Muliple Access wih Collision Avoidance (). I uses a Binary Exponenial Backoff (BEB) echnique o randomise each conender aemp of ransmiing, effecively reducing he collision probabiliy. Neverheless, relies on a random backoff ha while effecive and fully decenralised, in principle is unable o compleely eliminae collisions, herefore degrading he nework hroughpu as more conenders aemp o share he channel. To overcome hese siuaions, Carrier Sense Muliple Access wih Enhanced Collision Avoidance (CSMA/ECA) is able o creae a collision-free schedule in a fully decenralised manner using a deerminisic backoff afer successful ransmissions. Hyseresis and Fair Share are wo exensions of CSMA/ECA o suppor a large number of conenders in a collision-free schedule. CSMA/ECA offers beer hroughpu han and shorerm hroughpu fairness. This work describes CSMA/ECA and is exensions. Addiionally, i provides he firs evaluaion resuls of CSMA/ECA wih non-sauraed raffic, channel errors, and is performance when coexising wih nodes. Furhermore, i describes he effecs of imperfec clocks over CSMA/ECA and presen a mechanism o leverage he impac of channel errors and he addiion/wihdrawal of nodes over collision-free schedules. Finally, experimenal resuls on hroughpu and los frames from a CSMA/ECA implemenaion using commercial hardware and open-source firmware are presened. Index Terms CSMA/ECA, WLAN, MAC, Collision-free, Tesbed. I. INTRODUCTION Wireless Local Area Neworks (WLANs or WiFi neworks []) are a popular soluion for wireless conneciviy, wheher in public places, work environmens or a home. This echnology works over an unlicensed specrum in he Indusrial, Scienific and Medical (ISM) radio bands (a around 2.4 or 5 GHz), offering a good radeoff beween performance and coss, which is a main reason for is populariy. The Medium Access Conrol (MAC) scheme used in WLANs is based on Carrier Sense Muliple Access wih Collision Avoidance () proocol. I has been widely adoped by manufacurers and consumers, making i inexpensive o implemen and an ubiquious echnology. Neverheless, he ever-growing hroughpu demands from upper layers are This work was parially suppored by he Spanish governmen, hrough he projec CISNETS (TEC ). L. Sanabria-Russo (luis.sanabria@upf.edu), J. Barcelo (jaume.barcelo@upf.edu) and B. Bellala (boris.bellala@upf.edu) are wih he Deparmen of Informaion and Communicaion Technologies, Universia Pompeu Fabra, Barcelona, Spain. F. Gringoli (francesco.gringoli@ing.unibs.i) is wih he Diparameno di Eleronica per l Auomazione, Universià degli Sudi di Brescia, Brescia, Ialy. faced wih a boleneck a he WLANs MAC [2], which by is naure is prone o collisions ha degrade he overall performance as more nodes join he nework [3]. The research communiy has pushed forward many alernaives o he curren MAC in WLANs [4] [4], bu when a proposal deviaes oo much from, or some imecriical operaions are modified, is hardware implemenaion as par of WLANs MAC ofen becomes unlikely [5], wih he sandardisaion process aking many years wihou cerainy of approval [2]. A replacemen should be able o provide advanages in erms of hroughpu, specrum efficiency and number of suppored conenders. All of he aforemenioned wihou sacrificing shor-erm hroughpu fairness. Furhermore, WLANs implemening such a replacemen mus also serve exising users, which means hey have o be backwards compaible. A suiable candidae, and he one o be evaluaed in his work, is called Carrier Sense Muliple Access wih Enhanced Collision Avoidance (CSMA/ECA) [8]. I is capable of aaining higher hroughpu han by making a simple modificaion o he conenion mechanism, hus keeping mainaining compaibiliy. In CSMA/ECA, nodes use a deerminisic backoff afer successful ransmissions, consrucing a collision-free schedule among successful conenders in a fully decenralised way. This backoff mechanism ensures ha more channel ime is spen on successful ransmissions raher han recovering from collisions, hus increasing he hroughpu of he nework. Furher enhancemens (or exensions), like Hyseresis and Fair Share [] allow CSMA/ECA o suppor many more conenders in a collision-free schedule. Moreover, CSMA/ECA and is exensions are designed for allowing nodes o ransmi as frequenly as possible while keeping an even disribuion of he available bandwidh among users. The 82. High Efficiency WLAN (HEW) Task Group (TG) envisions very crowded scenarios as one of he fuure challenges for WLAN proocols [], [8], specifically hose usually encounered a sadiums or conference rooms. Furher, if he need for serving many users is combined wih he increased hroughpu demands from he upper layers, a performance improvemen a he MAC becomes paramoun. Alhough many sudies have been made analysing he performance of CSMA/ECA [8], [9], [], [9], here are sill several open aspecs ha require furher aenion and addiional insigh o consider CSMA/ECA as a poenial replacemen. Namely, neiher assesses he proocol s backwards compaibiliy propery or is behavior under non-sauraed raffic condiions while serving many users. Furhermore,

2 2 he impac of channel errors, node addiion/wihdrawal, and imperfec clocks over he deerminisic backoff mechanism is also lacking. This paper fills hose gaps by exending [], and consolidaes he push for CSMA/ECA as a poenial replacemen of for nex generaion WLANs. In deail, his paper provides he following conribuions: Firs resuls on he achievable hroughpu and delay of CSMA/ECA wih Hyseresis and Fair Share under nonsauraed raffic condiions for very large number of nodes. The impac of imperfec clocks and channel errors on CSMA/ECA wih Hyseresis and Fair Share s deerminisic backoff and is consequences on he achieved performance. Formulaion of he hroughpu bounds for CSMA/ECA wih Hyseresis and Fair Share. Inroduce he Schedule Rese mechanism for reducing he schedule lengh in case of users wihdrawing from he conenion (non-sauraed raffic), or when is lengh is increased due o channel errors. Coexisence and backwards compaibiliy wih nodes under differen raffic condiions. Firs implemenaion of CSMA/ECA wih Hyseresis in real hardware and he experimenal resuls in a crowded WLAN esbed. Resuls, derived from a modified version of he COST [] simulaor show ha CSMA/ECA wih Hyseresis and Fair Share is capable of accommodaing many users in collisionfree schedules. As ess in a mixed nework wih differen proporions of and CSMA/ECA nodes show, he aggregaed hroughpu is higher han he observed in -only neworks. Furhermore, a low number of oal conenders s hroughpu is acually improved, while i is degraded in crowded scenarios. This consiues a moivaion for a change owards CSMA/ECA. Beyond simulaions resuls, he implemenaion of CSMA/ECA prooypes [2] [24] show ha he consrucion of collision-free schedules using a deerminisic backoff afer successful ransmissions is possible and resuls in a hroughpu increase by reducing he number of corruped frames. We hen presen he firs real hardware implemenaion of CSMA/ECA wih Hyseresis using he open firmware OpenFWWF [25]. Resuls show ha CSMA/ECA is able of providing higher hroughpu and lower losses han for he same number of conenders. An overview of similar decenralised and collision-free MAC proocols for WLANs is provided in Secion II. CSMA/ECA, as well as is exensions for allocaing many conenders in a collision-free schedule are explained in Secion III. Secion IV deails he simulaion environmen for esing CSMA/ECA, while Secion V explains he resuls. An overview of CSMA/ECA real-hardware implemenaion and esbed descripion is compiled in Secion VI, followed by a summary of he sill missing CSMA/ECA feaures needed o become he nex replacemen in Secion VII. Conclusions are drawn in Secion VIII. II. RELATED WORK Time in WLANs is divided ino iny empy slos of fixed lengh σ e, collisions, and successful slos of lengh σ c and σ s, respecively. Collision and successful slos conain collisions or successful ransmissions, making hem several orders of magniude larger han empy slos (σ e min(σ s, σ c )). One of he effecs of collisions is he degradaion of he nework performance by wasing channel ime on collisions slos. Recen advances in he WLANs PHY [2], [2] push he research communiy owards he developmen of MAC proocols able o ake advanage of a much faser PHY. By reducing he ime spen in collisions nodes are able o ransmi more ofen, which in urn ranslaes o an increase in he nework hroughpu. Furher, he upcoming MAC proocols for WLANs should work wihou message exchange beween conenders, ha is, work in a fully decenralised fashion in order o avoid injecing exra conrol raffic ha may reduce he daa hroughpu. Performing ime slo reservaion for each ransmission is a well known echnique for increasing he hroughpu and mananining Qualiy of Service (QoS) in TDMA schemes, like LTE [2]. Applying he same concep o CSMA neworks by modifying DCF s random backoff proceedure provides similar benefis []. The following are MAC proocols for WLANs, decenralised and capable of aaining greaer hroughpu han by consrucing collision-free schedules using reservaion echniques. A survey of collision-free MAC proocols for WLANs is presened in [28]. In his paper we only overview hose ha are similar o CSMA/ECA. A. Zero Collision MAC Zero Collision MAC (ZC-MAC) [29] achieves a zero collision schedule for WLANs in a fully decenralised way. I does so by allowing conenders o reserve one empy slo from a predefined virual schedule of M-slos in lengh. Backlogged saions pick a slo in he virual cycle o aemp ransmission. If wo or more saions picked he same slo in he cycle, heir ransmissions will evenually collide. This forces he involved conenders o randomly and uniformly selec oher empy slo from hose deeced empy in he previous cycle plus he slo where hey collided. When all N saions reserve a differen slo, a collision-free schedule is achieved. ZC-MAC is able o ouperform under differen scenarios. Neverheless, given ha he lengh of ZC MAC s virual cycle has o be predefined wihou acual knowledge of he real number of conenders in he deploymen, he proocol is unable o provide a collision-free schedule when N > M. Furhermore, if M is overesimaed (M N), he fixed-widh empy slos beween each conender s successful ransmission are no longer negligible and conribue o he degradaion of he nework performance. Addiionally, ZC- MAC nodes require common knowledge of where he virual schedule sars/ends. This is no considered in and consiues an obsacle owards sandardisaion. B. Learning-MAC Learning-MAC [28] is anoher MAC proocol able o build a collision-free schedule for many conenders. I does so

3 3 defining a learning srengh parameer, β (, ). Each conender sars by picking a slo s for ransmission of he schedule n of lengh C a random wih uniform probabiliy. Afer a conender picks slo s(n), is selecion in he nex schedule, s(n + ), will be condiioned by he resul of he curren aemp. () and (2) exraced from [28] show he probabiliy of selecing he same slo s(n) in cycle n +. } p s(n) (n + ) =, Success () p j (n + ) =, p s(n) (n + ) = βp s(n) (n), p j (n + ) = βp j (n) + β C, Collision (2) for all j s(n), j {,..., C}. Tha is, if a saion successfully ransmied in s(n), i will pick he same slo on he nex schedule wih probabiliy one. Oherwise, i follows (2). The selecion of β implies a compromise beween fairness and convergence speed, which he auhors deermined β =.95 o provide saisfacory resuls. L-MAC is able o achieve higher hroughpu han wih a very fas convergence speed. Neverheless, he choice of β suppose a previous knowledge of he number of empy slos (C N, where N is he number of conenders), which is no easily available o or may require a cenralised eniy [4]. Furher exensions o L-MAC inroduced an Adapaive schedule lengh in order o increase he number of suppored conenders in a collision-free schedule. This adapive schedule lengh is doubled or halved depending on he presence of collisions or many empy slos per schedule, respecively. The effecs of reducing he schedule lengh may provoke a reconvergence phase which can resul in shor-erm fairness issues. Moreover, L-MAC nodes also require common knowledge of he sar/end of he schedule. III. CARRIER SENSE MULTIPLE ACCESS WITH ENHANCED COLLISION AVOIDANCE (CSMA/ECA) CSMA/ECA [8] is a fully decenralised and collisionfree MAC for WLANs. I differs from in ha i uses a deerminisic backoff, B d = CW min /2 afer successful ransmissions, where CW min is he minimum conenion window of ypical value CW min =. By doing so, conenders ha successfully ransmied on schedule n, will ransmi wihou colliding wih oher successful nodes in fuure cycles []. Collisions are handled as in, which is described in Algorihm. In Algorihm, he node sars by seing he reransmissions couner and backoff sage o zero (r [, R] and k [, m] respecively, wih m he maximum backoff sage, and R = m + he reransmissions limi. The ypical value for m is 5), hen generaes a random backoff, B. When he Acknowledgemen (ack) for a sen packe is no received by he sender a collision is assumed. Upon collision, he involved nodes will double heir conenion window by incremening heir backoff sage in one and use a random backoff, B [, 2 k CW min ]. This procedure is described beween Line and 3 of Algorihm. Algorihm 2 provides an explanaion of CSMA/ECA s deerminisic backoff mechanism, which main difference wih (and herefore wih Algorihm ) relies on he selecion of a deerminisic backoff afer a successful ransmission (compare Line 8 in Algorihm wih Line 8 in Algorihm 2). Figure shows an example of CSMA/ECA dynamics wih four conenders. while he device is on do 2 r ; k ; 3 B U[, 2 k CW min ]; 4 while here is a packe o ransmi do 5 repea while B > do wai slo; 8 B B ; 9 Aemp ransmission of packe; if collision hen r r + ; 2 k min(k +, m); 3 B U[, 2 k CW min ]; 4 unil (r = R) or (success); 5 r ; k ; if success hen 8 B U[, 2 k CW min ]; 9 else Discard packe; 2 B U[, 2 k CW min ]; 22 Wai unil here is a packe o ransmi; Algorihm :. r indicaes he number of reransmission aemps, while R is he maximum reransmission aemps limi. When R reransmissions are reached, he packe waiing for ransmission is dropped. In Figure, he STA-# labels represen saions willing o ransmi. The horizonal lines represen a ime axis wih each number indicaing he amoun of empy slos lef for he backoff o expire. Saions willing o ransmi begin he conenion for he channel by waiing a random backoff, B. The firs ouline highlighs he fac ha saions STA-3 and STA-4 will evenually collide because hey have seleced he same B. Afer recompuing he random backoff, STA-4 s aemp resuls in a successful ransmission, which insrucs he node o use a deerminisic backoff, B d = in his case. By doing so, all successful STAs will no collide among each oher in fuure cycles. Collision slos being orders of magniude larger han empy slos degrade he nework performance. When CSMA/ECA builds he collision-free schedule all conenders are able o successfully ransmi more ofen, increasing he aggregaed hroughpu beyond s. Figure 2, shows he average aggregaed hroughpu of CSMA/ECA, and

4 4 STA STA STA STA Fig.. An example of he emporal evoluion of CSMA/ECA in sauraion. Slos conaining ransmissions are around imes larger han empy slos while he device is on do 2 r ; k ; 3 B U[, 2 k CW min ]; 4 while here is a packe o ransmi do 5 repea while B > do wai slo; 8 B B ; 9 Aemp ransmission of packe; if collision hen r r + ; 2 k min(k +, m); 3 B U[, 2 k CW min ]; 4 unil (r = R) or (success); 5 r ; k ; if success hen 8 B d 2 k CW min /2 ; 9 B B d ; else 2 Discard packe; 22 B U[, 2 k CW min ]; 23 Wai unil here is a packe o ransmi; Algorihm 2: CSMA/ECA. collision-free proocolos presened in Secion II, namely ZC- MAC and L-MAC [28]. Referring o Figure 2, CSMA/ECA is able o achieve an aggregaed hroughpu ha goes beyond up unil he number of conenders is greaer han B d + (N = 8 in he case of he figure). Beyond his poin, he nework will have a mixed behaviour relaing o backoff mechanisms: some nodes will successfully ransmi and use a deerminisic backoff while ohers will collide due o he lack of empy slos and reurn o a random backoff. As more conenders join he nework, CSMA/ECA performance will approximae o s. This effec is also observed in ZC-MAC and L-MAC. I happens becasue he virual cicle used by hese proocolos (M for ZC-MAC and C for L-MAC in Secion II) Aggregaed Throughpu (Mbps) 5 CSMA/ECA Hys CSMA/ECA ZC-MAC L-MAC JFIs below JFI for JFI for CSMA/ECA JFI for ZC-MAC JFI for L-MAC JFI for JFI for CSMA/ECA Hys Fig. 2. CSMA/ECA example in sauraion: hroughpu (simulaion parameers can be found in Secion IV) in lower han he number of conenders, N. The JFI curves in Figure 2 show he Jain s Fairness index for all esed proocols. Showing a JFI equal o one suggess ha he available hroughpu is shared evenly among all saions. A. Supporing many more conenders As was menioned before, CSMA/ECA is only able o build a collision-free schedule if he number of conenders N, is less or equal han B d +. When N > B d +, collisions reappear. To be able o aain a collision-free schedule even when he number of conenders exceeds B d +, we inroduce Hyseresis. Hyseresis is a propery of he proocol ha insrucs nodes no o rese heir backoff sage (k) afer successful ransmissions, bu o use a deerminisic backoff B d = CW (k)/2, where CW (k) = 2 k CW min. This measure allows he adapaion of he schedule lengh, admiing many more conenders in a collision-free schedule. This idea of a schedule is significanly differen from he virual schedule required by he proocols described in Secion II, ha is, CSMA/ECA wih Hyseresis does no require a previous knowledge of he number of conenders, he resul of previous ransmissions or L-MAC and ZC-MAC curves in Figure 2 appear o yield he same aggregaed hroughpu. This is because hese proocols do no consider an increase in he backoff sage afer a failed ransmission, augmening he collision probabiliy beyond s. An adapaion of he schedule lengh leverages his issue, which is preseed in [28] as Adapive L-MAC and L-ZC. Jain s Fairness Index (JFI)

5 5 Noaion k m B B d CW min CSMA/ECA Hys CSMA/ECA Hys+MaxAg FS CSMA/ECA MaxAg TABLE I CSMA/ECA NOTATION Descripion Backoff sage Maximum backoff sage Random backoff Deerminisic backoff Minimum Conenion Window CSMA/ECA wih Hyseresis CSMA/ECA wih Hyseresis and Fair Share CSMA/ECA wih Hyseresis and Maximum Aggregaion wih Fair Share wih Maximum Aggregaion he sar/end of he schedule, easing is implemenaion in real hardware. Hyseresis enables CSMA/ECA nodes o have differen schedules (B d ), carrying he undesired effec of unevenly dividing he channel ime among conenders (i.e., some nodes will have o wai more in order o aemp ransmissions). This unfairness issue is solved by insrucing nodes a backoff sage k o ransmi 2 k packes on each aemp, hus proporionally compensaing hose nodes a higher backoff sages. This addiional exension o CSMA/ECA is called Fair Share. CSMA/ECA wih Hyseresis and Fair Share will be referred o as in order o disinguish i from wha was described unil his poin. The idea of allowing he ransmission of more packes o saions ha ransmi less ofen was iniially proposed by Fang e al. in [28]. I was laer adaped o CSMA/ECA Hys and named Fair Share in []. Figure 2 shows he JFI for as well as for. In Figure 2, he only curve deviaing from JFI = is CSMA/ECA wih Hyseresis (CSMA/ECA Hys ), suggesing an uneven pariion of he channel access ime among conenders (which is fixed wih Fair Share). Algorihm 3 describes, Table I provides a shor lis of noaions used hroughou he ex, while an example of wih four conenders is shown in Figure 3. In he figure he firs ouline indicaes a collision beween STA-3 and STA-4, which will provoke an incremen on boh saions backoff sage (k k + ). Once STA- 4 s random backoff expires, insrucs he saion o ransmi 2 k packes, and hen use a deerminisic backoff, B d = CW (k)/2. The same procedure is followed by STA 3. Wih Hyseresis and Fair Share, is able o achieve greaer hroughpu han and for many more conenders, as shown also in Figure 2. In he figure, he curve shows a greaer hroughpu because collisions are eliminaed and Fair Share allows nodes o send 2 k packes upon each ransmission. This hroughpu increase is he resul of aggregaion via Fair Share. I carries he negaive effec of raising he average ime beween successful ransmissions (see Secion V-A3), which may affec delaysensiive raffic, like gaming or live video/voice/v sreaming. Channel errors, hidden-nodes, or unsauraed raffic condiions will disrup any collision-free schedule, generae colli- while he device is on do 2 r ; k ; k c k; 3 b U[, 2 k CW min ]; 4 while here is a packe o ransmi do 5 repea while B > do wai slo; 8 B B ; 9 Aemp ransmission of 2 k packes; if collision hen r r + ; 2 k min(k +, m); 3 B U[, 2 k CW min ]; 4 unil (r = R) or (success); 5 r ; if success hen B d 2 k CW min /2 ; 8 B B d ; 9 else Discard 2 kc packes; 2 B U[, 2 k CW min ]; 22 k c k; 23 Wai unil here is a packe o ransmi; Algorihm 3: : k c refers o he conenion backoff sage, ha is, he backoff sage wih which a conenion for ransmission is sared. Afer R reransmission aemps, Fair Share insrucs he node o drop 2 kc packes. sions and push all saions backoff sages o he maximum value very quickly (conribuing o he delay increase). This issue is leveraged wih he Schedule Rese mechanim, inroduced in Secion III-G. B. The effecs of Aggregaion Fair Share is an A-MPDU aggregaion mechanism [] ha coupled wih he collision-free schedule buil by CSMA/ECA Hys is able o provide shor-erm fairness. However, i also improves he hroughpu since he aggregaion process makes he packe ransmission more efficien by reducing overheads. Furhermore, he level of aggregaion provided by Fair Share depends only on he buffer occupancy and he curren backoff sage. The downside of Fair Share is ha i may increase he ime beween wo consecuive ransmissions from he same node, which may affec negaively delay-sensiive applicaions such as gaming or high definiion real-ime video. An example of he duraion of a ransmission, T (l k ), is provided by (5) in he following Secion III-C. In scenarios where shor-erm fairness and he ime beween consecuive ransmissions are no relevan, Fair Share can be replaced by Maximum Aggregaion (MaxAg), which will significanly improve he sysem hroughpu. In Maximum Aggregaion all nodes aggregae as many packes as possible a every ransmission, i.e., hey send 2 m packes in each aemp.

6 STA STA STA STA Fig. 3. An example of he emporal evoluion of in sauraion (CW min = ) Aggregaed Throughpu (Mbps) 5 CSMA/ECA Hys+MaxAg MaxAg FS JFIs below JFI CSMA/ECA Hys+MaxAg JFI MaxAg JFI JFI FS JFI Fig. 4. Throughpu comparison wih MaxAg : even hough a low number of conenders MaxAg achieves greaer hroughpu han, collisions evenually degrade he hroughpu below s when he number of conenders increases pas N = Figure 4 shows he aggregaed hroughpu for, CSMA/ECA Hys+MaxAg, using he Fair Share mechanism ( FS ), and MaxAg. JFIs are shown below. Alhough FS and MaxAg perform aggregaion, collisions degrade he aggregaed hroughpu as more conenders aemp ransmission. On he oher hand, is able o build a collision-free schedule and akes advanage of he aggregaion provided by Fair Share, which opposed o jus using CSMA/ECA Hys+MaxAg, i is fair. To summarise he effecs of using aggregaion: I increases he aggregaed hroughpu: because nodes are able o send muliple packe in each aemp, he sysem hroughpu is increased. Moreover, Fair Share compensaes hose nodes a higher backoff sages o ensure hroughpu fairness. Maximum aggregaion supposes he deacivaion of he Fair Share mechanism: performing maximum aggregaion upon each ransmission aemp is equivalen o having differen schedule lenghs and no compensaing nodes a higher backoff sages. Alhough he aggregaed hroughpu increases, his resuls in an uneven disribuion of he channel ime among conenders, which renders i unfair. Longer periods beween ransmission aemps: given ha each ransmission akes longer, he ime beween Jain s Fairness Index (JFI) ransmission aemps also increases. This may specially affec delay-sensiive applicaions. Since we consider ha fairness and a shor inerransmission ime are even more imporan han raw hroughpu for he nex generaion of WLANs, we keep as he reference proocol. C. Throughpu bounds of They correspond o he maximum and minimum achievable hroughpu wihou he possibiliy of collisions using Hyseresis and Fair Share. The ideal nework uses he minimum schedule lengh ha guaranees a collision-free operaion. Tha is, wih a schedule lengh of C = 2 k B d +, where k = log 2 (N/(B d + )). Using his minimum schedule lengh, N nodes will be a he same backoff sage if N B d +. Oherwise, h = N (C N) nodes would occupy he k-h backoff sage and he oher N h nodes he (k )-h one. The sysem hroughpu is compued as follows: S = { hs k (l k ) + (N h)s k (l k ), if N > B d + Ns k (l k ), oherwise where s k (l k ) and s k (l k ) are he hroughpu achieved by he nodes a he k-h and (k )-h backoff sages sending l k and l k packes respecively. These are given by: l k L s k (l k ) = ht (l k ) + 2(N h)t (l k ) + σ e (C N) l k L s k (l k ) = (N h)t (l k ) + k T (l k) 2 (3) (4a) (4b) where L is he daa payload, T (l k ) and T (l k ) are he duraion of he ransmission of l k and l k packes, respecively; σ e is he duraion of an empy slo. Addiionally, T (l k ) derives from (5): ( ) SF + lk (MD + MH + L) + TB T (l k ) = T PHY + T sym L ( DBPS ) SF + LBA + TB + SIFS + T PHY + T sym + DIFS + σ e L DBPS (5) where T PHY = 32 µs is he duraion of he PHY-layer preamble and headers, T sym = 4 µs is he duraion of an

7 Aggregaed Throughpu (Mbps) Fig. 5. Maximum Aggregaion (Hys+MaxAg) Upperbound (Hys+FS) Lower-bound (Hys+FS) CSMA/ECA Upper and Lower hroughpu bounds for D. Clock drif issue in descenralized collision-free MAC proocols CSMA/ECA relies on saions being able o correcly coun empy slos and consequenly aemp ransmissions in he appropriae slo according o he backoff imer. Failure o do so may be caused by clock imperfecions inside he Wireless Nework Inerface Cards (WNIC), which is commonly referred o as clock drif. As poined ou in [32], clock drif is a common issue ha degrades he hroughpu in disribued collision-free MAC proocols like he ones reviewed in Secion II. While miscouning empy slos have no significan effec on s hroughpu [32], i has a direc impac on CSMA/ECA. In a collision-free schedule wih sauraed CSMA/ECA conenders, a saion miscouning empy slos will drif o a possibly busy slo, collide and force a reconvergence (if possible) o a collision-free schedule (see Secion V-A2). OFDM (Orhogonal Frequency Division Muliplexing) symbol. SF is he Service Field ( bis), MD is he MPDU Delimier (32 bis), MH is he MAC header (288 bis), TB is he number of Tail Bis ( bis), L BA is he Block-ACK lengh (25 bis) and L DBPS = 25 is he number of bis in each OFDM symbol. SIFS, DIFS and σ e values can be found in Table II. The Lower-bound is derived from considering he operaion of an ideal nework. Nodes use he minimum backoff sage possible and aggregae proporionally, hus yielding he minimum hroughpu achievable by a nework. I is compued following (3) wih l k = 2 k and l k = 2 k. The Upper-bound is obained from considering he operaion of a nework using, bu forcing nodes o use he maximum backoff sage for deermining he cycle lengh and he level of aggregaion. I is also compued using (3) bu considering ha all nodes are in he maximum backoff sage (k = m) and herefore l k = 2 m. The maximum hroughpu achievable is he resul of deacivaing he Fair Share rules by forcing nodes o use maximum aggregaion regardless of heir backoff sage. This is called Maximum Aggregaion (Hys+MaxAg) in Figure 5. I can be derived from (3) considering l k = 2 m and l k = 2 m. I is ineresing o see in Figure 5 how collisions force colliding conenders o increase heir backoff sage and aggregae more wih Fair Share. This explains why he curve separaes iself from he Lowerbound a a very low number of conenders. Alhough using maximum aggregaion (see Maximum Aggregaion (Hys+MaxAg) curve in Figure 5) increases he hroughpu i carries he effec of unevenly disribuing he available bandwidh among conenders, as menioned in Secion III-B. The ools for deriving hese wo curves are available as MATLAB funcions in [3]. E. Channel errors Failed ransmissions due o channel errors are handled as collisions by. Therefore, collision-free periods under his ype of channel model are frequenly inerruped. In order o accelerae he convergence o collisionfree schedules in presence of channel errors insruc nodes o sick o he curren deerminisic backoff even afer sickiness number of consecuive failed ransmissions. Sickiness has been inroduced o CSMA/ECA in [4], where i is described and evaluaed. Sickiness allows for a faser convergence owards a collision-free schedule, especially when performing under heavy channel errors. Failed ransmissions due o channel errors also means ha a few momens of operaion under a noisy channel can increase he conenders s deerminisic backoff o is maximum value. Such even carries he undesired effecs of increasing he ime beween successful ransmissions and reducing he overall sysem hroughpu due o fewer collision-free periods of operaion. F. Hidden erminals One key characerisic of IEEE 82. devices is ha heir carrier sense range is a leas wo imes greaer han heir daa range [33]. In his siuaion, he impac of hidden nodes can be considered o be low. This is because a given ransmission could only be inerfered by oher ransmissions from very disan nodes wih energy levels no higher han he noise floor, which using modern radios can be easily solved by he capure effec. However, in specific deploymens, where obsacles play also an imporan role on he propagaion effecs, hidden nodes may appear, causing asynchronous packe collisions [34]. In he case of WLANs, he same negaive effecs as in IEEE 82. WLANs are expeced. In addiion, collisions wih hidden erminals will cause nodes o leave any collision-free schedule, as channel errors do, hus coninuously disruping any aemp of collision-free operaion.

8 Bd = Transmi, hen change: Bd * = Fig.. Example of he Schedule Rese mechanism. The lower row shows values of for a bimap b[] of size b[] = B d + = 32 G. Schedule Rese Mechanism insrucs nodes o keep heir curren backoff sage afer a successful ransmission (reseing i o zero only if he node empies is MAC queue, see line 2 in Algorihm 3). This is done in order o increase he cycle lengh and provide a collision-free schedule for many conenders, which is desirable in dense scenarios. Neverheless, having a big deerminisic backoff increases he ime beween successful ransmissions. If no operaing in a scenario wih many nodes, he empy slos beween ransmissions are no longer negligible and degrade he overall hroughpu of he sysem. For insance, if nodes wihdraw from he conenion (i.e.: empy heir MAC queue, or move o oher nework) heir previously used slos will now be empy. On he oher hand, in scenarios wih channel errors conenders rapidly end up wih he larges deerminisic backoff, remaining here unil he MAC queue empies. Nodes should be aware of his issue and pursue opporuniies o reduce heir deerminisic backoff, B d wihou sacrificing oo much in collisions. The Schedule Rese mechanism for consiss on finding he smalles collision-free schedule (if any) beween a conender s ransmissions and hen change he node s deerminisic backoff o fi in ha schedule. Take a conender wih B d = 3 as an example. By lisening o he slos beween is ransmissions, i is possible o deermine he availabiliy of smaller (and possibly) collision-free schedules. Figure shows he slos beween he ransmissions of a conender wih B d = 3. Saring from he lef, he curren B d = means ha he firs slo will be filled wih he conender s own ransmission. Each following slo conaining eiher a ransmission or a collision is idenified wih he number one, while empy slos are marked wih a zero. Noice ha he highlighed empy slos appear every eigh slos, suggesing ha a schedule reducion from B d = 3 o Bd = is possible; where B d is he new deerminisic backoff assigned by Schedule Rese 2. The schedule change is performed afer he conender s nex successful ransmission. ) A conservaive schedule reducion: Schedule Rese is described in Algorihm 4. Each conender sars filling a bimap b of size B d + wih he informaion observed in each passing slo during γ = C/B d consecuive successful ransmissions (sxtx in Algorihm 4); where C = 2 m CW min /2 is he bigges deerminisic backoff, and γ is referred o as he Schedule Rese hreshold from his poin forward. 2 B d = 2 CW min /2 = (see Table II). So he change is made o he backoff sage, k. In his case from 2, o k =. This means ha any new schedule mus also be a power of wo. if sxtx == hen 2 Iniialising; 3 if B == hen 4 b[b d + ] = {}; 5 = ; if (sxtx > ) && (sxtx γ) hen Filling he bimap; 8 b[] σ i (); 9 + +; if == B d hen = ; 2 if sxtx == γ hen 3 Analising; 4 sxtx = ; 5 for (j = ; j < k; j + +) do y = 2 j CW min /2 ; isipossible = ; 8 for (m = y; m < B d + ; m+ = y) do 9 isipossible + = b[m]; if isipossible == hen 2 change = ; 22 newsage = m; 23 break; 24 if change == hen 25 Making he change; 2 k = newsage; 2 change = ; Algorihm 4: Schedule Rese Mechanism for. Every consecuive successful ransmission increases he variable SxTx by one, while a collision reses i o zero. The algorihm is called a rese when all smaller deerminisic backoff are esed. On he oher hand, when j in line 5 is iniialised o j = k, i is called a schedule halving Each bi, {,..., B d } in he bimap is he resul of a biwise OR operaion beween is curren value and he sae of he corresponding slo. This is shown in Line 8 in Algorihm 4, where σ i () is a funcion ha evaluaes o if he slo corresponding o b[] is empy in ieraion i, i {,..., γ}; or oherwise. Afer γ ieraions, he bimap is evaluaed (line 2 in Algorihm 4). Schedule Rese ess all possible deerminisic backoffs ha are smaller han he curren B d (lines 5-9 in Algorihm 4), saring wih he smalles one. If he corresponding bis in he bimap are regisered as empy, he process is sopped and he change o he new deerminisic backoff is made (lines -2 in Algorihm 4). Oherwise, he process is resared afer he nex successful ransmission. Using Figure as an example, we can see ha b[] =, for {8,, 24}. This means ha he ransmission slos corresponding o a deerminisic backoff, Bd = are free, herefore a change of schedule is possible. In case of suffering

9 9 a collision immediaely afer applying he schedule reducion, he node reverses he changes made by Schedule Rese before handling he collision. 2) Aggressive approach o face channel errors: he defaul γ ensures ha he bimap regisers all he ransmission slos in he nework (assuming sauraed raffic and a perfec channel), providing enough informaion for performing he schedule reducion wihou disruping collision-free schedules. Neverheless, his γ can be rendered oo conservaive and unnecessary because nodes randomly leave he collision-free schedule due o errors. This condiion produces Schedule Rese bimaps ha do no conain enough informaion for successfully avoiding collisions afer he schedule reducion. This is also rue for any γ > γ. This effec is leveraged by seing γ =, meaning ha he bimap is filled wih he informaion of slos beween only wo consecuive ransmissions. This measure increases he frequency of schedule rese aemps. Furhermore, incremening he node s sickiness afer a successful reducion of he schedule allows for a faser convergence owards a collision-free schedule, especially when performing under heavy channel errors. H. Backwards compaibily and coexisence springs from a modificaion o s backoff mechanism. I keeps he range of values nodes use o draw a random backoff (i.e., use he same CW min and CW max ), allowing conenders o coexis wih nodes in he same nework. Furher, he selecion of s deerminisic backoff, B d, is he expeced value for he curren backoff sage k (B d := E[, CW (k) ] ) [], which ensures fairness among conenders. An overview of he aained hroughpu for differen proporions of and nodes is presened in Secion V-C. IV. SIMULATION SCENARIO This secion provides he simulaion parameers for esing under wo differen raffic condiions, namely sauraed and non-sauraed. We also provide deails on how channel errors are modelled and wha are is effecs over he ransmissions. Furher, he simulaion of he clock drif effec, Schedule Rese parameers, and he coexisence wih are also subjecs o be addressed in his secion. A. Scenario deails Resuls are obained by running muliple simulaions over a modified version of he COST simulaor [], available a [35]. PHY and MAC parameers are deailed in Table II. Some assumpions were made in order o es he performance a he MAC layer: Unspecified parameers follow he IEEE 82.n (2.4 GHz) amendmen. All nodes are in communicaion range. No Reques-o-Send (RTS) or Clear-o-Send (CTS) messages are used. TABLE II PHY AND MAC PARAMETERS FOR THE SIMULATIONS PHY Parameer Value PHY rae 5 Mbps Empy slo 9 µs DIFS 28 µs SIFS µs MAC Parameer Value Maximum backoff sage (m) 5 Minium Conenion Window (CW min ) Maximum reransmission aemps Daa payload (Byes) 24 MAC queue size (Packes) Collisions ake as much channel ime as successful ransmissions. The aforemenioned assumpions ensure ha he simulaion resuls are jus effecs of he MAC behaviour. If no menioned oherwise, resuls are derived from simulaions of seconds in lengh, each one wih a differen seed. Figures also show he sandard deviaion. B. Sauraed and Non-sauraed saions A sauraed saion always has packes in is MAC queue. This is modelled by seing he packe arrival rae o he MAC queue ( PAR ) o a value greaer han he achievable hroughpu. To ensure sauraion, saions are se o fill heir MAC queue a PAR = 5 Mbps, which is purposefully greaer han he effecive capaciy of he channel. To evaluae he performance under non-sauraed condiions, saions need o be able o empy heir MAC queues. To do so, he packe arrival rae o he MAC queue is se o PAR = Mbps. These values of PAR have proven o produce he desired effecs. C. Performance under clock drif Clock drif is simulaed by seing a drif probabiliy, p cd. Each saion has a probabiliy of p cd /2 of miscouning one slo more, and p cd /2 of miscouning one slo less. This approach follows he one proposed by Gong e. al in [32]. D. Channel errors Channel errors are modeled by assigning a probabiliy of a packe being corruped by he channel, p e >. Tha is, in a single packe ransmission here is probabiliy p e ha he ransmission will no be acknowledged. If he ransmission is an A-MPDU (like in he case of ), p e will affec each MPDU individually and independenly. Therefore, an A-MPDU ransmission will be considered a complee failure only if all frames in he aggregaion are affeced by he channel error probabiliy. All resuls are shown wih sickiness equal o one (see Secion III-E). Tha is, afer a colision, nodes will use a random backofff. Neverheless, in Secion V-D curves described as dynsick emporarily increase he node s sickiness o wo afer a successful reducion of he schedule

10 (using a random backoff afer wo consecuive collisions). This increase is done in order o converge faser o collision-free schedules when operaing wih channel errors. E. Coexisance wih To es he performance of and saions in he same nework, simulaions are se wih a node densiy of /4, /2 and 3/4 of he oal. F. Applying Schedule Rese A se of resuls under sauraed condiions and channel errors applying Schedule Rese (see Secion III-G) are provided. Some of he resuls are generaed wih a γ =, or aggressive Schedule Rese (indicaed as aggr. in he figures). These seings provide he highes hroughpu under he esed condiions, and also in he real hardware implemenaions such as he one shown in Secion VI. A. Sauraed nodes V. RESULTS In, a large number of sauraed nodes will normally be relaed o a high collision probabiliy. This effec is in par he resul of reseing he backoff sage afer a successful ransmission and he generaion of a new random backoff. However, his scenario provides an advanageous condiion o nodes. In sauraion, nodes build a collision-free schedule and sick o heir deerminisic backoff as long as hey ransmi successfully, effecively eliminaing collisions. This secion aims a overviewing he hroughpu of and in sauraion, as well as he collision probabiliy, he average ime beween successful ransmissions and he effec of clock drif over he hroughpu. ) Throughpu: nodes are able o build a collision-free schedule, use he channel more efficienly, and experience a hroughpu increase as seen in Figure a. Hyseresis allows he allocaion of more conenders by increasing he lengh of he collision-free schedule, while Fair Share ensures an even disribuion of he available hroughpu. This is refleced by he average backoff sage, which value increases wih he number of conenders. In conras, hroughpu keeps decreasing due o an augmened number of collisions as he number of nodes increases (see Figure c). Furher, Figure b shows he fracion of collision slos for and as simulaion ime passes. In he figure, is appreciaed how he fracion of collision slos keeps decreasing once reaches collisionfree operaion. 2) Effec of clock drif over he achieved hroughpu in sauraion: Figure e shows he nework aggregaed hroughpu wih sauraed saions and an increasing clock drif probabiliy. In Figure e, a saion has a clock drif probabiliy equal o p cd. Each saion has a probabiliy of p cd /2 of miscouning one slo more, and p cd /2 of miscouning one slo less. Because is based on a random backoff, miscouning slos has no significan effec on he hroughpu. For he CSMA/ECA curve, i is possible o appreciae a sligh decrease of he hroughpu as p cd increases, caused by collisions due o he drif. The curve in Figure e shows insead an increase of he aggregaed hroughpu as p cd grows. Collisions make conenders o incremen heir backoff sage and aggregae packes for ransmissions according o Fair Share, effecively increasing he hroughpu. As i also can be appreciaed in he figure, he average backoff sage for conenders increases rapidly o is maximum value (m = 5), reducing he effec of clock drif over nodes given ha heir ransmissions would now be separaed by more slos. 3) Time beween successful ransmissions: I is relaed o he elapsed ime beween he conenion for ransmission and he recepion of an ACK. In Figure d all ess wih maximum aggregaion, namely MaxAg and CSMA/ECA Hys+MaxAg, have an increased average ime beween successful ransmissions. This is due o he muliple packes ha are sen in each aemp. MaxAg, hough, has an increased value due o collisions also aking longer channel ime. Alhough has an increased average ime beween successful ransmissions due o Fair Share, i has a lower meric when compared wih he maximum aggregaion curves in Figure d. B. Non-sauraed nodes Empying he MAC queue in means ha nodes will rese heir backoff sage o zero and use a random backoff when a new packe arrives a he queue, breaking any collision-free operaion in. The following show he impac over hroughpu, delay and ime beween successful ransmissions when using and in non-sauraed condiions. ) Throughpu: In Figure 8a, he aggregaed hroughpu increases linearly for he curve unil sauraion is reached a around 22 nodes, where he hroughpu begins o degrade. The curve has a similar behavior, enering sauraion a around nodes. Furher, a around nodes we see an increase in he average backoff sage for conenders which suggess an incremen in collisions. This effec is shown in Figure 8b and Figure 8d, where a around 35 nodes conenders sar colliding and dropping packes. As indicaed by Figure 8b, when < N 35 nodes suffer from an increasing number of collisions. This is due o nodes empying heir MAC queue quicker due o Fair Share, as shown in Figure 8f and reenering he conenion wih a random backoff every ime a new packe arrives. This increase in collisions may also cause he dropping of packes due o reaching he maximum reransmission limi.

11 a) b) e) Aggregaed Throughpu (Mbps) Average % of collision slos BOS for BOS for Backoff sages below 5 c) 5 Average backoff sage (BOS) Fracion of collision slos... e+ 4e+4 8e+4 Avg. Time be. sx x (secs) Time slos d) MaxAg CSMA/ECA Hys+MaxAg Aggregaed Throughpu (Mbps) 5 CSMA/ECA BOS for BOS for CSMA/ECA BOS for Backoff sages below Average backoff sage (BOS) Clock drif probabiliy, p cd (%) Fig.. Simulaion resuls under sauraed raffic: a) Throughpu under sauraed condiions; b) Evoluion of he fracion of collision slos in a scenario wih sauraed saions; c) Average percenage of collision slos: fracion of ime slos conaining collisions; d) Average ime beween successful ransmissions (sx x), e) Throughpu when increasing he clock drif probabiliy As drops more packes afer reaching such limi (see line in Algorihm 3), i shows a higher fracion of dropped packes in Figure 8d. Beyond 35 conenders, he MAC queue of nodes sars o fill up, gradually allowing longer periods of collision-free operaion due o nodes geing sauraed. This allows o ouperform. 2) Delay: This meric refers o he elapsed ime beween he injecion of a packe ino he saion s MAC queue and he recepion of an ACK for such packe. In Figure 8e, a rapid increase in he delay for nodes is appreciaed a he sauraion poin (around conenders), whereas s delay is sill low. Furher, wih he percenage of blocked packes from he MAC queue is lower han or MaxAg (see Figure 9). This is due o he consrucion of collision-free schedules which ensure ha large A-MPDU ransmissions do no suffer from collisions. As nodes ge sauraed, he delay increases due o longer queueing and conenion ime (see he number of packes in he MAC queue for nodes in Figure 8f and how i is relaed o he increase in delay shown in Figure 8e). Figure 8c shows he average ime beween successful ransmissions. I is possible o see from he figure ha when approaches he sauraion poin he average ime beween successful ransmissions increases, resembling Figure d. Average percenage of blocked packes (%) non-sauraed MaxAg non-sauraed CSMA/ECA Hys+MaxAg non-sauraed non-sauraed 5 Fig. 9. Average fracion of blocked packes. When he MAC queue is full, packes comming from he uppers are discarded (or blocked) a he enrance of he buffer C. Coexisence wih CSMA/ECA is hough o be an evoluion of given is similariies and he abiliy o coexiss wih he laer. This secion provides simulaions resuls for a seup of differen proporions of nodes in a nework where here are also conenders, ha is: /4, /2 and 3/4 of he oal nodes run, while he res uses. This nework configuraion will be referred o as mixed nework seup from here on. ) Throughpu: Figure a shows he nework hroughpu for differen proporions of nodes in a mixed nework seup. In he figure i is appreciaed how he mixed nework seups

12 2 Fracion of dropped packes (%) Aggregaed Throughpu (Mbps) 9e-4 8e-4 e-4 e-4 5e-4 4e-4 3e-4 2e-4 e-4 e+ a) BOS for : BOS for Backoff sages below 5 d) 5 5 Average backoff sage Average (%) of collision slos Average Delay (secs) b) 5 e) MaxAg (MaxAg.) 5 Avg. num. of packes in MAC queue Avg. Time be. sx x (secs) 8e- e- e- 5e- 4e- 3e- 2e- e- e+ e+3 8e+2 e+2 4e+2 2e+2 e+ c) MaxAg CSMA/ECA Hys+MaxAg 5 f) 5 Fig. 8. Simulaion resuls under non-sauraed raffic: a) Throughpu; b) Average percenage of collision slos: he fracion of ime slos conaining collisions; c) Average ime beween successful ransmissions; d) Average fracion of dropped packes; e) Average sysem delay; f) Average number of packes in he MAC queue of a node curves lay beween he and curves. As he proporion of nodes decreases, he hroughpu increases as he resul of a lower probabiliy of collision, as can be seen in Figure b. A similar behaviour is observed when esing he same proporion of nodes under non-sauraed condiions. Figure c and Figure d show he average aggregaed hroughpu and fracion of collisions slos in a non-sauraed mixed nework seup. As shown in Figure a and Figure c, a a lower proporion of nodes (/4) he average aggregaed hroughpu is he greaes among he esed mixed nework seups. This is because collisions rigger Hyseresis and Fair Share in nodes, lowering he number of imes hese nodes ener in a conenion and reducing he overall collision probabiliy when compared o an only nework (see Figure b, Figure d and Figure e). As he proporion of nodes increases, he nework hroughpu approximaes o. Figure e shows he average hroughpu for and saions in a sauraed mixed seup (half he conenders using ). I is possible o see ha for a low oal number conenders (N 2) saions aain greaer hroughpu han in a -only nework. Again, his is because in he mixed nework seup he oher N /2 conenders wih use a deerminisic backoff, leaving many empy slos beween ransmissions. Sill referring o Figure e, for N > 2 periods of collision-free operaion and he aggregaion performed wih Fair Share allows o have larger channel ime han nodes, which hroughpu degrades even more han in -only. These resuls suggess ha a swich o can be beneficial for neworks wih high number of conenders. D. Channel errors and Schedule Rese Figure a shows he average aggregaed hroughpu of a sauraed nework wih a channel error probabiliy, p e =. (see Secion IV-D). We can see ha he highes hroughpu, corresponding o is lower han he one observed in a perfec channel (Figure a). Furhermore, curves wih Schedule Rese (SR) seem o have lower aggregaed hroughpu. This is because he backoff sage, and herefore he level of aggregaion done by Fair Share, is repeaedly reduced. Sill assessing Figure a and focusing on he SR curves, we can see ha boh aggressive schedule halving wih dynamic sickiness (SR aggr. halv. dynsick) curves (wih and wihou Fair Share) have higher aggregaed hroughpu han he res (for a descripion of aggressive schedule halving and dynamic sickiness, refer o Algorihm 4 and Secion III-G2). This Schedule Rese configuraion reduces he ime beween successful ransmissions (see Figure c) and mainains collisionfree operaion for longer periods of ime hanks o he increase in he node s sickiness. Figure c shows he average ime beween successful ransmissions for he same nework seup. Even-hough sill has a higher meric han due o he aggregaion done by Fair Share, Schedule Rese is able of reducing his meric by almos 43%. Figure b and Figure d show he aggregaed hroughpu and ime beween successful ransmissions, respecively for a non-sauraed nework wih he same p e =.. Since under non-sauraed condiions nodes rese heir schedule every ime hey empy heir MAC queues, Schedule Rese s benefis are no relevan. Neverheless, a reducion