1 Analyss of Collaboratve Dstrbuted Admsson Control n 82.11x Networks Thnh Nguyen, Member, IEEE, Ken Nguyen, Member, IEEE, Lnha He, Member, IEEE, Abstract Wth the recent surge of wreless home networks, t s ncreasngly mportant to employ admsson control mechansms n order to enhance the performance of the wreless multmeda applcatons. In ths paper, we extended the work n [1] for performng dstrbuted admsson control n a collaboratve wreless envronment. In partcular, we provde an n-depth analyss on the throughput and jtter of traffc of the proposed admsson control algorthms for collaboratve wreless networks. Smulaton results confrm our theoretcal predctons on the performance of the proposed admsson control algorthms. I. INTRODUCTION Recent years have wtnessed an explosve growth n multmeda wreless applcatons such as vdeo streamng and conferencng. One of the reasons for ths tremendous growth s the deployment of IEEE 82.11x WLANs [2] n both prvate homes and enterprse networks. Wthout a proper bandwdth allocaton mechansm n a wreless network, a flow may experence a sgnfcant performance degradaton due to a large number of flows. To allevate ths problem, the recent 82.11e standard [3][4][5] specfes a mechansm for traffc dfferentaton. However, traffc dfferentaton alone cannot guarantee the performances for ndvdual flows sharng the wreless medum. Instead, admsson control mechansm s needed to ensure that a new flow s admtted, only f there s enough resource to support the requred throughputs of all the exstng flows. Unlke the admsson control n a wred network, admsson control for a wreless network s more complex. Ths complcaton s due to the channel contenton access, n whch the nterferences.e., the collsons between the new flow and the exstng flows reduce all the flow s throughputs. The number of these collsons ncreases nonlnearly wth the number of competng flows, makng t harder for an admsson control algorthm to decde whether or not to admt a new flow. Many exstng admsson control algorthms for 82.11x based networks have been proposed [6][7][8]. Most of them assume an Access Pont (AP) that runs the admsson control algorthm. The advantage of ths approach s that the AP has complete nformaton about the network, and thus can allocate the resource effcently. For example, Xao and L [6][9] uses the measurements to provde flow protecton (solaton) n 82.11e network. The algorthm requres the AP to broadcast the necessary nformaton to other wreless statons. On the other hand, centralzed admsson control s not possble when multple wreless devces form an ad-hoc network. Thus, the devces must collaborate wth each other to jontly satsfy ther bandwdth requrements. In ths paper, we descrbe and analyze the performance of a collaboratve dstrbuted admsson control (CDAC) algorthm n whch, every devce are assumed to be honest and collaboratve [1]. Specfcally, a devce wll not start a new flow f by dong so, there s not suffcent resources to support the exstng flows. Our dscusson wll be lmted to a one-hop wreless network, e.g. the network of all the wreless devces wthn a home or a small buldng such Thnh Nguyen and Ken Nguyen are wth the Oregon State Unversty. Lnha He s wth Boston Consultng Group.
2 that, every devce can hear the transmssons of all other devces. We frst dscuss the background and a few related work n Secton II. Next, we descrbe the proposed CDAC algorthm and ts analyss n Secton III. Secton IV s devoted to the dscusson of the smulaton results. Fnally, we summarze our paper n Secton V. A. Contenton Based Access n 82.11x Networks II. PRELIMINARIES Contenton based access enables multple devces to compete for a shared wreless channel. Whle there are many parameters n 82.11x standards, for smplcty, we focus our dscusson on the contenton wndow (CW) and TXOP. To access the channel, a devce frst senses the channel. If the channel s dle for more than the Arbtraton Interframe Space (AIFS) tme, the devce starts sendng the data. Otherwse, t sets a back-off tmer for a random number of tme slots between [,CW mn ] where CW mn s the mnmum contenton wndow sze. The backoff tmer s decremented by one for each dle tme slot after the AIFS tme, and halts decrementng when a transmsson s detected. The decrementng resumes when the channel s sensed dle agan for an AIFS tme. A devce can begn transmsson on the channel as soon as ts backoff tmer reaches zero. If a collson occurs,.e., no acknowledgment packet s receved after a short perod of tme, the back-off tmer s chosen randomly between [, (CW mn + 1)2 1] where s the number of retransmsson attempts. Clearly, a flow uses a small CW mn s lkely to obtan the channel than a flow that uses a large CW mn. One problem wth contenton based access n the 82.11 WLAN s the nonlnear throughput reducton of the exstng flows wth the ntroducton of a new flow due to hgher collson rate. To that end, Pong and Moors [1] devses an algorthm for computng the new CW mn s and TXOP s for the exstng flows to mantan ther requred throughputs. Ther algorthm s based on the analytcal model proposed by Banch [11]. In ths model, the transmsson probablty p t () of a flow can be derved wth the followng formula: 2(1 2p c ()) p t () = (1 2p c ())(W +1)+p c ()W (1 (2p c ()) b ), (1) Where p c () s the long term collson probablty of flow, W s the CW mn used for flow, and b s the maxmum backoff stage wth CW max =(CW mn + 1)2 b 1. Pong and Moors propose to montor the collson probablty p c () of each flow n order to calculate transmsson probablty p t. Once p t s obtaned, one can compute the probablty of a successful slot p s () foraflow. The throughput of a flow then can be computed usng p s (). Please see [1] for the detal calculatons. The PM algorthm makes an approxmaton by assumng that the collson probablty of a flow remans constant before and after the new flow jons. Another dffculty wth the PM algorthm s that when a new flow requests to jon, the AP has to approxmate the collson rate of a new flow as the collson rate of an exstng flow wth smlar throughput. If a flow wth smlar throughput s not avalable, the algorthm may produce naccurate throughputs for all the flows. III. COLLABORATIVE DISTRIBUTED ADMISSION CONTROL (CDAC) Our algorthm s smlar to the PM algorthm n the sense that every devce also adjusts ts transmsson parameters to mantan ther requred throughputs n presence of the new flow. However there are several major dfferences
3 as follows. Frst, unlke the PM algorthm, we do not assume that the collson probablty of every flow remans constant before and after the new flow jons. Therefore, our algorthm can accurately calculate the throughput for each flow. Second, guessng the collson probablty of a new flow s not requred n our algorthm. Thrd, our algorthm s desgned to be performed dstrbutedly whle the PM algorthm must be done at the AP. Fnally, unlke the PM algorthm, our algorthm assumes that the contenton wndow sze s not doubled after every unsuccessful retransmsson attempt. We argue that doublng of contenton wndow sze s not optmal when a proper admsson control s used. Usng admsson control, the traffc load at any gven tme s known or hghly predctable, and therefore the contenton wndow sze and/or the TXOP of each devce can be precsely tuned (by the admsson control algorthm) to acheve the requred throughput. Therefore, we expect hgher performance from ths approach than that of blndly doublng the contenton wndow sze after every unsuccessful retransmsson attempt. Furthermore, we consder a small wreless network n whch every wreless devce can hear each other s transmssons such as a network of all the wreless devces (wthout an AP) wthn a home. We also assume the use of reservaton packets RT S/CT S. In our proposed scheme, each wreless devce contnuously montors traffc n ts neghborhood and estmates the followng parameters: I: percentage of dle slots (denoted as type-i slots). S : percentage of slots wth a successful transmsson of RTS/CTS packet (denoted as type-s slots) of flow. C: percentage of slots wth collson (denoted as type-c slots). D : throughput of flow as a fracton of the channel capacty. Note that I + C + S =1. Also, f admsson control s performed at the AP, these parameters can be easly obtaned. Our analyss s based on tme-slotted, reservaton based protocols, where the tme taken to make a reservaton s a geometrcally dstrbuted random varable wth parameter p. To translate the transmsson probablty p back to the contenton wndow sze used n IEEE 82.11x protocol, W can be set to 1/p. We note that ths s only an approxmaton snce W n IEEE 82.11x protocols are not reset at every tme slot. Suppose a devces wants to start a new flow of throughput x kbps. It frst needs to decde whether the network has suffcent resources to support the new flow. If t does, how would each devce tune ts transmsson parameters, e.g. CW, to take nto account of the new traffc n order to mantan the requred throughputs. Each devce s assumed to be collaboratve such that t wll not nject a new flow f there s not enough bandwdth for the exstng flows. It s mportant to note that all the admsson control operatons are done by the devce that wants to nject a new flow. If and only f the new flow s admtted by the admsson control algorthm runnng locally on the same devce, then ths devce wll broadcast the updated transmsson parameters to other devces. Ths approach reduces the workload for other devces. In ths paper, we only consder tunng the contenton wndow sze. A. Theoretcal Performance of N flows We assume that every devce can start at most one flow at any pont n tme. A straghtforward generalzaton to support multple flows per devce s to consder all the flows from one devce as one sngle large flow wth the transmsson probablty p. Whenever a devce successfully obtans the channel, t selects a packet from one of ts
4 flows to send. The probablty of a packet selected from a partcular flow then equals to the rato of that flow s throughput to the total throughput of all the flows on the same devce. Ths approach would result n the correct average requred throughputs for all the ndvdual flows. We now show how to compute I, C, and S s, gven the transmsson probabltes p s. Suppose the transmsson probablty for a new flow s p, then for the type-c slots, n whch collsons occur, the new traffc would have no mpact on them. For the type-s slots, wth probablty p, t may cause a collson n that slot. For any type-i slot, wth probablty p, t would become a type-s slot. Otherwse t stays the same. Usng the above argument, we can calculate the percentages of C-type, I-type, and S-type slots after the new flow starts. In partcular, the new dle, collson, and successful slots can be calculated usng the current I, C, S, and p as: I new = I current I current p (2) C new = C current + S current p (3) S new = S current (1 p)+i current p. (4) Here, we denote S = S. Smlarly, we can calculate the percentages of successful slots S as S,new = S,current (1 p), (5) for any exstng flow, and the percentage of successful slots for the new flow as S N = I current p. (6) Usng the equatons above, t s straghtforward to derve an algorthm for computng I, C, S s for N users n terms of the transmsson probabltes p 1,p 2,...p N teratvely. Due to lmted space, we omt the algorthm s lstng and smply call ths algorthm: Algorthm 1. In prncple, the Algorthm 1 for computng S s n terms of p s produces a set of N equatons wth N unknown varables p 1, p 2,... p N. Hence, one can solve for p s based on S s. Unfortunately, these equatons are not lnear, and therefore dffcult to solve. We propose an algorthm to fnd the p s gven S s based on the followng observaton. We note that when a flow stops, I wll ncrease by S.Ifflows starts agan wth the same transmsson probablty p as before, ts percentage of successful slots remans S as before. Hence, the followng equatons hold: (I + S )p = S ; p = S. (7) I + S Ths s true because I +S s the percentage of dle slots wthout flow. Hence, after the flow starts, ts percentage of successful slots s (I + S )p whch should also precsely equals to S, the successful percentage before t stops. Thus, we have N such equatons correspondng to N flows. We also have another equaton: I + C + S =1, (8) where C and S are the percentages of collson and successful slots for all the flows. We note that I s the same for every equaton snce t s the percentage of dle slots when all flows are actve. Now, we can solve for N +1 unknowns,.e. N p s and I. Solvng ths set of N +1 equatons s smple snce each equaton s lnear except
5 Equaton (8). However, Equaton (8) wll be used as a constrant. The dea s to systematcally try dfferent values of I from small to large, e.g. to 1. For each value of I, we compute p s accordng to Equaton (7). All the p s are then nput the Algorthm 1 to compute C and S. We then test to see whether or not I + C + S approxmately equals to 1. If so, we have an admssble set of solutons. If not, we ncrease I by a small value and repeat the procedure. If the algorthm cannot fnd satsfed p s for the entre range of I [, 1], then ths mples nvald S s. Due to lmted space, we omt the algorthm s lstng, and smply call ths algorthm: Algorthm 2. To verfy our theoretcal results, we smulate the followng scenaro. Fve flows start n the ncreasng order - wth flows 1 and 5 beng the frst and the last n the network (note that the startng order of flows does not change the results). The transmsson probabltes of flows 1 to 5 are.5,, 5,, and 5, respectvely. The TXOP for all the flows s set to 1 tme slots and remans constant throughout the smulaton. Fgure 1(a) shows the percentages of dfferent slot types as a functon of the number of flows for both theory and smulaton. As seen, the theory and smulaton results match each other almost exactly for dfferent types of slots. As the number of flows ncreases, the collson and successful probabltes ncreases whle the dle probablty decreases. Ths result s plausble when the network load s stll lght. In other words, every flow succeeds n gettng ther slots from a large number of dle slots. Fgure 1(b) shows the percentages of successful slots for dfferent flows decreases as the number of flows ncreases as expected. Fnally, Fgure 1(c) shows the percentages of the total bandwdth taken by dfferent types of slots as the number of flows ncreases. Percentage of dfferent types of slots 1.9.8.7.6.5.4 success smulaton success theory dle smulaton dle theory collson smulaton collson theory 1 1.5 2 2.5 3 3.5 4 4.5 5 Number of flows Percentage of successful slots for each flow 6 4 2.8.6.4 flow 4 flow 5.2 1 1.5 2 2.5 3 3.5 4 4.5 5 Number of flows Percentage of bandwdth for each type of slots.9.8.7.6.5.4 success RTS bandwdth dle bandwdth collson bandwdth data bandwdth 1 1.5 2 2.5 3 3.5 4 4.5 5 Number of flows (a) (b) (c) Fg. 1. Fve flows jons the network sequentally, each wth dfferent transmsson probabltes; (a) The percentages of successful, collson, and dle slots vs. the number of flows from theory and smulaton; (b) Percentages of successful slots of all flows as the number of flows ncreases; (c) Percentage of bandwdth taken by dfferent types of slots. B. Admsson Control Algorthm Assume a fxed TXOP, we proceed to calculate the throughput n terms of the number of tme slots per unt tme. The throughput n kbps can be easly obtaned by multplyng the tme slot throughput by the number of bts per tme slot. In addton, we assume that every devce knows the maxmum capacty of the channel. Thus, all the calculatons of throughput wll be done n terms of a fracton of the maxmum channel capacty. It s straghtforward to compute S n terms of TXOP and the fractonal throughput D s - the requested throughput as a fracton of the channel capacty as: D S = S (1 N j=1 D (9) j)
6 Next, usng Algorthm 2, p s can be computed gven S s. The admsson control algorthm (shown below) wll accept the new connecton only f there exsts a set of vald p s. Algorthm 3: Admsson Control Ad_Ctrl_p(D[], TXOP, N) { D = sum of all D[] s; for = 1:1:N { S[] = D[]/(TXOP(1-D)); } // computng p s. If algorthm 2 cannot fnd a set of p_, the value of success s zero. [p[], success] = Algorthm_2(S[], N); f success = { new flow s rejected. } else { admt the new flow. } } When a devce requests to start a new flow, t uses the Algorthm 2 to determne whether or not a set of p s exsts. If so, t broadcasts the updated p s to other devces, and starts sendng ts data wth ts transmsson probablty calculated from Algorthm 2. Other the devces can start transmttng ther flow wth the new p s. Otherwse, f no p s exst, the new flow s rejected. C. Throughput Jtter Analyss The proposed admsson control algorthm guarantee that each flow wll obtan ts requred throughput when t s averaged over a long perod of tme. However, the throughputs of the flows may fluctuate wthn a short perod of tme. These fluctuatons n throughput prevent smooth playback for many audo and vdeo streamng applcatons. Thus, we would lke to analyze the throughput jtter for dfferent flows wthn a short perod of tme. A devce can estmate the throughput jtters of exstng flows and use them as addtonal requrements to decde whether or not to nject ts new flow. We proceed to quantfy the throughput jtter of a flow as the normalzed standard of devaton of ts fractonal throughput wthn a number of tme slots. Specfcally, t s: Stdev n (X (T )) = Stdev(X (T )), (1) D Where X (T ) s a fracton of the data slots of the flow measured wthn T tme slots, and D s ts average long term fractonal throughput. Clearly, f the throughput s measured wthn small wndow of tme (number of tme slots), ts varance wll ncrease. Now, the amount of data sent by a flow wthn a number of tme slots s proportonal to the number of successful transmssons of the RTS/CTS wthn that perod of tme. Thus, we can calculate the varances of the number of successful transmsson slots wthn a perod of tme. The probablty of successful transmsson for a flow for a slot s S. The number of non-data slots n T tme slots s T (1 N D ) where D s the fractonal throughput of flow. Thus the number of successful transmsson slots Y wthn ths perod s bnomally dstrbuted wth the parameters S and T (1 N D ). Therefore, ts varance s: N Var(Y )=T(1 D )S (1 S ). (11) By defnton, X (T )=Y TXOP /T, (12)
7 Therefore, the varance of X (T ) s: Var(X (T )) = (1 N D )S (1 S )TXOP 2 T, (13) and the standard of devaton of X (T ) s (1 N D )S (1 S ) Stdev(X (T )) = TXOP. (14) T Fnally, the normalzed standard of devaton of the fractonal throughput equals to: (1 N D )S (1 S ) Stdev n (X (T )) = TXOP (15) D T IV. SIMULATION RESULTS AND DISCUSSION To evaluate the performance of our admsson control algorthms, we set up the followng smulaton scenaro. Two flows 1 and 2 are assumed to send data contnuously whle requests to send data at dfferent rates. Flow 3 uses the proposed admsson control algorthm to fnd out whether or not there s enough bandwdth for t to send data. If so, t determnes the parameters for the exstng flows 1 and 2, and sends these parameters to all the flows. The transmsson probabltes for flows 1 and 2 are set.5 and.2, respectvely. The TXOP s of all the flows are set to 1 tme slots, and reman constant throughout the smulatons. As a result, D 1 =9 and D 2 =2. Flow 3 requests dfferent throughputs, rangng from.5 to.4 of the total capacty. Fgure 2(a) shows the new transmsson probabltes to mantan the requred throughputs for all three flows. As requests a larger porton of the throughput capacty, the transmsson probabltes of all three flows ncrease to compensate for smaller probabltes of successfully obtanng the channel. Fgure 2(b) shows the rato of the measured throughput to the requested throughput for each flow. Fgure 2(c) shows the correspondng measured throughputs. As seen, the algorthm s able to mantan the throughputs of the exstng flows whle provdes the requred throughput for the new flow. Fnally, 2(d) shows the reduced throughputs of the flows f no admsson control s used. We now show the throughput jtter for the proposed algorthm. The requested fractonal throughputs D s are set to 5, 2, and.5, respectvely. The TXOP s set to 1 tme slots. S s are then calculated n terms of D s and fxed TXOP. The normalzed standard of devaton of the fractonal throughputs are then computed usng Equaton (15). Fgure 3(a) shows the normalzed standard of devaton as a functon of the buffer sze. As expected, as the buffer sze ncreases, the normalzed standards of devaton decreases. Fgure 3(b) shows the normalzed standard of devatons of throughputs as a functon of the throughput of. As seen, the normalzed stdev. s of flows 1 and 2 reman constant whle the normalzed stdev s of flows 3 reduces as ts throughput ncreases. V. CONCLUSIONS In summary, we extended the work n [1] for performng dstrbuted admsson control n a collaboratve wreless envronment. In partcular, we provde an n-depth analyss on the throughput and jtter of traffc of the proposed admsson control algorthms for collaboratve wreless networks. Smulaton results confrm our theoretcal predctons on the performance of the proposed admsson control algorthms.
8 Updated transmsson probabltes 5 5 5.5.5 5 5 5.4 Bandwdth requested by Rato of measured to requested bandwdth 1.5 1.4 1.3 1.2 1.1 1.99.98.97.96.5 5 5 5.4 Bandwdth requested by Percentage of measured bandwdth.45.4 5 5 5.5 (a) Percentage of measured bandwdth.8.7.6.5.4 (b) cc.5 5 5 5.4 Bandwdth requested by (c).5 5 5 5.4 Bandwdth requested by (d) Fg. 2. (a) New transmsson probabltes as a functon of s bandwdth; (b) rato of measured to requested throughput as functon of s bandwdth; (c) flow s fractonal throughputs of flows 1 and 2 as a functon of s bandwdth; (d) flow s fractonal throughput wthout admsson control as a functon of throughput requested by. Normalzed standard devaton 5 5 5.5 Normalzed standard devaton.45.4.35.3.25.2.15 1 2 3 4 5 6 7 8 9 1 Number of tme slots for throughput averagng x 1 4 (a).1.5 5 5 5.4 Bandwdth requested by (b) Fg. 3. (a) Normalzed standard of devatons as a functon of (a) buffer sze and (b) throughput of ; REFERENCES [1] T. Nguyen, K. Nguyen, and L. He, Collaboratve dstrbuted admsson control for 82.11x networks, n Submtted to ICC, 27. [2] IEEE Std. 82.11, Wreless LAN Medum Access Control (MAC) and Physcal Layer (PHY) Specfcatons, 1999. [3] X. Yang, Ieee 82.11e qos provsonng at the mac layer, IEEE Wreless Communcaton Magazne, pp. 72 79, March 24. [4] IEEE Std. 82.11e/D6., Wreless LAN Medum Access Control (MAC) and Physcal Layer (PHY) Specfcatons: Medum Access Control (MAC) Qualty of Servce (QoS) Enhancements, 23. [5] S. Mangold and et.al, Analyss of eee 82.11e for qos support n wreless lans, IEEE Wreless Communcaton Magazne, pp. 4 5, March 23. [6] Y. Xao and H. L, Voce and vdeo transmssons wth global data parameter control for the eee 82.11e enhanced dstrbuted channel access, IEEE Transactons on parallel and dstrbuted systems, vol. 15, Novemeber 24. [7] YD. Gao, J. Ca, and K. Ngan, Admsson control n eee 82.11e wreless lans, IEEE Network, August 25. [8] D. Gu and J. Zhang, A new measurement-based admsson control method for eee 82.11 wreless local area networks, IEEE, 23. [9] Y. Xao, H. L, and S. Cho, Protecton and guarantee for voce and vdeo traffc n eee 82.11e wreless lans, IEEE, 24. [1] D. Pong and T. Moors, Call admsson control for eee 82.11 contenton access mechansm, n Globecom, December 23. [11] G. Banch, Performance analyss of the eee 82.11 dstrbuted coordnaton functon, IEEE JSAC, pp. 535 57, March 2.