A Model-based Scalable Reliable Multicast Transport Protocol for Satellite Networks

Transcription

1 24 JOURNAL O COMMUNICATION OTWARE AND YTEM, VOL., NO., EPTEMBER 2005 A Model-based calable Relable Multcast Transport Protocol for atellte Networks Prawt Chumchu, Roksana Borel, and Aruna enevratne Orgnal scentfc paper Abstract - In ths paper, we desgn a new scalable relable multcast transport protocol for satellte networks (RMT). Ths paper s the extensons of paper n [8]. The proposed protocol does not requre nspecton and/or ntercepton of packets at ntermedate nodes. The protocol would not requre any modfcaton of satelltes, whch could be bent-ppe satelltes or onboard processng satelltes. The proposed protocol s dvded n 2 parts: error control part and congeston control part. In error control part, we ntend to solve feedback mploson and mprove scalablty by usng a new hybrd of AR (Auto Repeat Request) and adaptve forward error correcton (AEC). The AEC algorthm adapts proactve redundancy levels followng the number of recevers and average packet loss rate. Ths leads to a number of transmssons and the number of feedback sgnals are vrtually ndependent of the number of recevers. Therefore, wreless lnk utlzaton used by the proposed protocol s vrtually ndependent of the number of multcast recevers. In congeston control part, the proposed protocol employs a new wndow-based congeston control scheme, whch s optmzed for satellte networks. To be far to the other traffcs, the congeston control mmcs congeston control n the wellknown Transmsson Control Protocol (TCP) whch reles on packet conservaton prncple. To reduce feedback mploson, only a few recevers, ACKers, are selected to report the recevng status. In addton, n order to avod drop-to-zero problem, we use a new smple wreless loss flter algorthm. Ths loss flter algorthm sgnfcantly reduces the probablty of the congeston wndow sze to be unnecessarly reduced because of common wreless losses. urthermore, to mprove achevable throughput, we employ slow start threshold adaptaton based on estmated bandwdth. The congeston control also deals wth varatons n network condtons by dynamcally electng ACKers. Index Terms Relable Multcast, atellte networks I. INTRODUCTION Multcastng greatly saves bandwdth and sender resources n applcatons, whch need to dstrbute nformaton to a selected group of end ponts. Internet Protocol (IP) provdes mechansms for provdng multcastng servces n networks whch are IP based, such as the Internet. Manuscrpt receved May 3, 2005; revsed June 28, 2005 and July 08, The paper was presented n part at the Conference on oftware, Telecommuncatons and Computer Networks (oftcom) Prawt Chumchu s wth Department of Telecommuncatons Engneerng, Mahanakorn Unversty of Technology, Bangkok 0530 Thaland. (Emal: prawt@mut.ac.th). A. enevratne and R. Borel are wth Natonal ICT Australa funded through the Australan Government's Backng Australa's Ablty ntatve, n part through the Australan Research Councl. However, t only provdes suffcent support for fxed hosts attached to wred terrestral networks. Moreover, ths support does not provde suffcent relablty. A number of applcatons need to dstrbute nformaton to many stes that are wdely dspersed from each other. atellte networks are deally suted for provdng the networkng servces for ths class of applcatons as they have several attractve characterstcs, such as breadth of broadcast reach, everywhere access, low-cost global coverage and flexblty capablty. It s also becomng ncreasngly clear that these servces are best provded usng IP. urthermore, there s a need to provde relable servces, as opposed smply best effort servces that are offered by IP. As a result, there has been a consderable effort at developng relable multcast transport protocols. However, most of the effort has focused on the desgn of relable multcast protocols have been for wred terrestral networks. These desgns have not taken nto account the nherent characterstcs of satellte channel and deployed network topologes, especally, the channel error characterstcs and long propagaton delays. In the lterature, several relable multcast protocols for terrestral networks such as n [2, 20] have been proposed. These protocols are not sutable for satellte networks because of dfferent envronments such as hgh delay, flat network, and hgh loss rate, etc. In addton, the congeston control algorthms such as n [6, 3, 2, 22] were desgned for wred networks, whch assume that all losses occur from congestons. Ths leads to congeston wndow or transmsson rate s unnecessarly decreased when losses occur from wreless lnks. Relable multcast transport protocols for satellte networks have been proposed n [5, 7,, 4, 7, 24]. Gouhong et al. [7] proposed a relable multcast protocol whch enables recevers to form clusters. Recevers of a cluster share to recover losses. Ths protocol does not provde congeston control algorthm. Chen et al. [24] proposed a relable multcast transport protocol for broadband satellte networks (RMTP). Ths protocol s not an end-to-end scheme, whch leads to a more complex mplementaton. Jan et al. [] have desgned a new relable multcast transport protocol. It s based on ACK-based scheme. In order to reduce feedback mploson and low bandwdth return lnks, they propose the logcal herarchcal groups. They assume some recevers mght have terrestral connectons whch could be used to create logcal groups. Basu et al. [5] mplemented a relable multcast protocol for hybrd satellte lnk. Error control part n ths protocol s based on a NAKbased protocol. Wth congeston control, they adopted a /05/ CCI

2 CHUMCHU et al.: A MODEL-BAED CALABLE RELIABLE MULTICAT TRANPORT PROTOCOL 25 montor-based flow control [22] whch assumes that all losses occur from congeston. Ths assumpton leads to face wth drop-to- zeros problem. MTP [7] s an end-to-end relable multcast transport protocol. The error control of MTP s a pass-based protocol. The pass-based protocol produces long transfer delay n the case of hgh loss rate and long delay channel. The congeston control s based on recever-controlled schemes whose recevers know ther capacty before jonng multcast group. The applcablty of multcastng protocols has to be consdered n vew of a specfc satellte network and applcatons of nterest, and, as demonstrated by the references, there s no generc one ft all soluton. In ths paper, we present a new Relable Multcast Transmsson protocol (RMT). Ths work has been performed as part of the Inmarsat multcastng project, and consequently the network model s based on the Inmarsat satellte system. However, the soluton s generc and can be appled to any satellte systems. The rest of the paper s structured as follows. ecton II presents the studed network model. ecton III presents overvew of consdered relable multcast transport protocols and the RMT framework. ecton IV presents the RMT error control part. ecton V presents the RMT congeston control part. ecton VI presents the extenson of the proposed protocol for multple satellte access nodes. The smulaton results are presented n secton VII and fnally secton VIII concludes and summarzes our work. rc ource RM Net u Congeston losses Common wreless losses patally ndependent wreless losses atellte Wred lnk rst mle lnk Last mle lnk Internet u atellte Access de (AN) AN Net LL/MAC Phy ErU ErD a. Drect-to-home deployment Channel atellte b. Network Model Channel g. Network Archtecture II. NETWORK MODEL Moble Earth taton (ME) R R2 ME RM Net LL/MAC Phy ErU ME RM Net LL/MAC Phy Network archtecture s shown n g. (a). Ths model s defned as the drect-to-home deployment n [4]. g. (b) shows the protocol stack model of the network archtecture. TABLE I defnes abbrevatons used n the network archtecture. It s composed by three network parttons: a wred lnk: the source and the AN (atellte Access de), a frst mle lnk: the AN and the satellte, and last mle lnks: the satellte and MEs (Moble Earth tatons). It could be analyzed that t s possble to be congested on only n the wred lnk. Ths s because n the frst mle lnk, the satellte ErU ErD ErD RM s desgned to forward the multcast packets. The forwardng rate has to be able to forward all ncomng traffc. In last mle lnks, t s assumed there s no congeston. Ths s because the sendng rate from the satellte s much less than MEs processng rate and also there s no sharng traffc from other network connectons. Packet losses could occur n all three parttons. However, packet error rate s very low for the wred lnk. If a packet s lost durng transmsson over the frst mle lnk, all recevers wll not receve the packet. Ths s called common wreless loss or spatally correlated wreless loss. In last mle lnks, t s assumed that the wreless loss s modeled as spatally uncorrelated wreless loss. Ths s because MEs locate n dfferent places. III. OVERVIEW O CONIDERED RELIABLE MULTICAT TRANPORT PROTOCOL AND RMT rstly we bref an overvew of operaton of MTP from tarburst [7]. In MTP, the sender dvdes a fle nto a number of Data Transmsson Unts (DTUs) called a block. MTP recommends that the block sze be ted to the Maxmum Transmsson Unts (MTU) sze so that the bt map whch exactly fts the number of DTUs n a block could be sent n one feedback packet. The transmsson rate of a block could be set at a gven level. MTP transmts the data by frst multcastng all data DTUs. Then, recevers report back what they have receved n a feedback packet, usng a smple bt map. The packets that have not been receved, are multcasted agan by the transmtter. Ths process contnues untl all recevers have successfully receved all the DTUs. Mechansm N3 s based on NAK-based schemes wth proactve EC. The frst transmsson of packet n ths scheme contans the orgnal packet as well as party packets whch are the same as n RMT s adaptve forward error correcton. If any group members who have not successfully receved the packet, are multcasted untl K rounds. After K rounds, any group members who have not successfully receved the packet, are uncasted untl all group members correctly receve the packet. The RMT framework could be classfed nto 2 parts, whch are shown n g. 2. The frst part s the error control part presented n secton IV, whch s based on a hybrd of AR, and Adaptve orward Error Correcton (AEC). The second part s the congeston control part presented n secton V. In the error control part, n order to reduce feed back mploson and mprove scalablty, the sender adapts the proactve redundancy level accordng to the number of actve recevers and estmated packet loss rate. The congeston control part reles on the packet-conservaton prncple [3]. The proposed congeston control uses a sngle-rate and wndow-based scheme. In order to reduce probablty of congeston wndow unnecessarly reduced because of spatally uncorrelated loss n last mle lnk, we use a smple loss flter algorthm (WLA). Ths loss flter algorthm uses a set of N recevers, (called ACKers) to report recevng status n form of ACK (Acknowledgement) messages. Each ACKer sends ACKs, whch contan the sequence number of newest receved data packet, a 32-bt btmap ndcatng the

3 26 JOURNAL O COMMUNICATION OTWARE AND YTEM, VOL., NO., EPTEMBER 2005 recevng status of the prevous 32 packets, echoed tme stamp, and estmated forward bandwdth. urthermore, we set slow start threshold and current congeston wndow based on estmated bandwdth when tme out or loss smlar to TCPW [8] whle standard TCP [3] sets slow start threshold to half of the current congeston wndow because the sender assumes that someone s sharng the network. In addton, our protocol deals wth ACKer changes and responses to varatons n network condtons such as ACKer falure. Abbrevatons RM Net LL/MAC Phy Er U Er D Er UM Er DM TABLE I Abbrevatons Meanng Relable Multcast Transport Protocols Network layer Data lnk layer: Logcal Lnk Layer and Meda Access Control layer Physcal Layer AN s Error module for return traffc AN s Error module for forward traffc ME s Error module for forward traffc ME s Error module for return traffc IV. ERROR CONTROL PART The operaton of error control part of RMT based N3 defned n [9] s descrbed as follows. The sender multcasts a block of n packets, whch conssts of k orgnal packets and h party packets. The selecton of h and k wll be presented n subsecton A. Multcast recevers check the packets they receve for errors through unque sequence numbers and unque EC block numbers. Ths s done ether durng the recepton of the next block or the sesson message f no data s beng receved. The sesson messages are perodcally sent n a new packet or attached a feld n data packets. If a multcast recever has only receved x dstnct packets and f x s less than k,.e. there are losses; t requests k-x new party packets. If a recever needs to repar a packet n the reported block, t sets a reschedule sendng NAK tmer. If the requested party packets are not receved before the tmer expres, the recever resends the NAK. Ths process contnues untl the recever successfully receves the requested packets. After recevng NAK, the sender sends new party packets or buffers for transmsson schedule to facltate the reparng of any lost packets. The transmsson rate s adapted followng the congeston control part descrbed n ecton V. A. Adaptve EC Algorthm n the error control part The proposed protocol uses the adaptve redundancy algorthm shown n g. 3. It takes the wreless loss parameters as nput, and calculates the number of requred redundancy packets by averagng the probablty of a multcast packet needng retransmsson after addng EC, q (. The value of q ( s calculated for the temporally uncorrelated loss model n (). The crtera parameter C determnes the number of redundancy packets, R h, to be retransmtted. The value of C R s R / R q(. The value of h s teratvely determned by ncreasng h untl the value of C R s less than X / R, where X s average packet loss rate of all recevers and R s the number of recevers. Let q( be the probablty that an orgnal packet s requested after EC recovery, gven by: k n n q ( = ( k ) p ( ( p( ), k = 0 n where p( s packet loss rate for ME j calculated n (2). B. Wreless loss parameter estmaton (LE) Each recever estmates the wreless loss parameters. A recever calculates the loss based on the gap of packet sequences or multcast sesson message. Therefore for each EC block, the recever calculates the probablty of loss; p b ( from the number of losses dvded by the number of packets n the EC block for the frst transmsson.e. excludng retransmtted packets. The average packet loss rate for recever j, p (, could be recursvely calculated usng the exponentally weghted movng average flter shown by followng equaton: b N B j= () p ( = ( z) p( + zp (,0 < z <, b = 0,,2... (2) where N B s the number of EC blocks and z s a smoothng factor, whch s greater than zero and less than one. In smulaton we set z to be 0.. Applcaton ueue3 EC Encoder LI Retransmsson Module ender ueue2 ueue Congeston sgnal(c) NAKs LE BWE A Transmsson Module Parameter Adjustment Algorthm ACKs g. 2 Relable Multcast Archtecture C. Negatve acknowledgements Recevers At the end of a block of EC (detected by recevng multcast sesson or by recevng a startng packet of a new EC block), a recever calculates the number of requested party packets for the EC block. To avod NAK mploson, the recever wats for a random tme, t, whch s randomly selected between 0 and t B before sendng the NAK packet to the sender. After sendng a NAK, the recever starts

4 CHUMCHU et al.: A MODEL-BAED CALABLE RELIABLE MULTICAT TRANPORT PROTOCOL 27 retransmsson tmer whch expres at next 2 RTTs (estmated round trp tme). Each NAK packet conssts of EC block ndex, the number of requested packet n the EC block, estmated packet loss rate, number of requested rounds for the EC block and echoed tme stamps plus t (for the frst requested round only). The echoed tme of frst requested round s used to estmate round trp tme by the sender. Intatng WP=0 W= ntates H T= AT=NULL Receved a block of k orgnal packets and p n = k+h Calculate q( C R < update h packets X M / ) /( top yes no h=h+ g. 3 the Adaptve EC Algorthm NL>0 WP = WP -WL H CAL n (6) W = H T = CA recevng an ACK WLA NA>0 WP= WP-NA T CA W = W+NA/W W>=H W = W+NA T=CA Tmeout WP=0 W= H CAL n (6) T = AT=NULL TCP protocols to adjust congeston control parameters. In our protocol, we use a wndow-based controller, whch wll be descrbed n next subsecton. A. Wndow Regulaton n Congeston Control Part The operaton of wndow regulaton n our proposed congeston control algorthm s shown n g. 4. The sender mantans the estmated number of packets outstandng n the path (W P ), the congeston control wndow sze (W), the token count (T), slow start threshold (H) and congeston control state (T). In our protocol, we assume that H and W are n unt of packets. The congeston control states could be dvded n to two states: low tart () and Congeston Avodance (CA). The operaton of the congeston control s presented as follows. After recevng an ACK, the sender performs loss flter algorthm (WLA) descrbed n next secton. The WLA gves the number of lost packets, N L, and the number of successfully receved packets, N A. If the number of loss packets s greater than 0, the sender calculates the slow start threshold (H) as n (3) and sets W to be H. Then t sets the congeston control state to Congeston Avodance. If the number of successfully receved packets s more than zero, the sender ncreases the wndow the same as AIMD (Addtve Increase Multplcatve Decrease). To avod stalls, the sender has tmer to restart probng the network after t could not send packets for a lmted tme. In our smulaton, we set the tme nterval to 2 RTTs. E xrtt M f E xrtt / M > 2 BW BW H = 2, otherwse / (3) Where E and RTT are estmated bandwdth and average BW round trp tme whch are descrbed n (5) and (6), respectvely. The varable M s the data packet sze n bts. T = W-WP end T packets, WP=WP+T B. Acknowledgements g. 4 Wndow Regulaton V. CONGETION CONTROL PART The proposed congeston control scheme tres to acheve fast response and elmnates the so-called drop-to-zero problem.e. the sesson rate droppng to value of wrong loss estmaton n the presence of uncorrelated losses. In addton, the scheme elmnates unnecessarly decrease n the multcast sesson rate when losses occur from wreless lnk errors. urthermore, the scheme stll retans scalablty because we use only a small number of recevers to be ACKers. In order to acheve these objectves, we desgn the proposed congeston scheme as follows. We use a smple loss ndcaton flter, whch could flter last mle losses out. The remanng losses are the losses occurred from wred lnk and the frst mle lnk. After usng the loss ndcaton flter algorthm, the communcaton could apply exstng wreless or each data packet, N recevers are n charge of sendng postve acknowledgement (ACKs). Each ACK contans the sequence number of the newest receved data packet, a 32-bt btmap ndcatng the recevng status of the prevous 32 packets, and echoed tme stamp. C. A New mple Wreless Loss Indcaton lter Three types of loss flter algorthms have been proposed n the lterature. The basc loss ndcaton s Worst Estmate- Based Trackng approach (WET) [2]. The second one s the Random Lstenng algorthm (RLA) [23]. The thrd one s the Lnear Proportonal Response approach (LPR) [6]. In [2] the sender elects a group representatve (Acker whose throughput s the worst). Ths algorthm s not sutable for satellte networks because of several reasons as follows. The worst recever mght change rapdly. The source unnecessarly reduces the congeston control wndow because of wreless lnk errors wth hgh probablty. To reduce these problems, we present a smple loss flter algorthm, Wreless

5 28 JOURNAL O COMMUNICATION OTWARE AND YTEM, VOL., NO., EPTEMBER 2005 loss Lstenng Algorthm (WLA) shown n g. 5. The objectve of the proposed loss flter algorthm s to pass only losses occurrng between the multcast sender and the satellte. The multcast sender dstngushes whether the packet losses occur between the multcast source and the satellte or the packet losses occur n the last mle lnk. Ths could be done by comparng the recevng status from all ACKers. If a packet s lost for all actve ACKers, the sender assumes that the loss occurred between the sender and the satellte. In order to avod packets out of sequences, A packet s assumed to be lost f the sender has not receved an ACK for the packet n a number of subsequence ACKs from all ACKers, dupack threshold. Ths value s set to 3 n our smulatons otherwse stated. D. Bandwdth Estmaton (BWE) In our congeston control, the sender sets slow start threshold followng estmated forward bandwdth smlar to TCPW [8]. The operaton of forward bandwdth estmate s descrbed as follows. The sender estmates the bandwdth usng the exponentally weghted movng average flter shown n r ( ( y) r ( yt ( =, (4) + where y s a constant between 0 and. The value t ( s M /( A A ), A s the tme that an ACK packet has been successfully receved and A s the tme that the prevous ACK packet was receved. After the sender could estmate average bandwdth as n (5) E = BW r ( N j where and N are the set of ACKers whose t are receved n predefned calculatng tme and the number of respectvely. Intalzng N T=NULL A T=NULL Lsq=- Receve ACK and 32-bt status Update A T and N T =Lsq+, NA=0 srt Ack NA=NA+ =+ >Msq NL=0,sp = NT LO N L=N L+ Delete NAK nformaton from NT g. 5 WLA Operaton E. Round Trp Tme Estmaton at the sender sp = sp->next, The average round tme estmaton of recever j, RTT (, could be recursvely calculated usng the exponentally weghted movng average flter shown by followng equaton. RTT ( ( g) RTT ( + grtt ( = (6) where RTT ( s sample round trp tme, g s a constant between 0 and. In smulaton we use g = 0.. The sample tme s calculated from echoed tme stamp sent by recevers n form of an ACK or a NAK. After predefned calculatng tme, the sender could estmate average round trp tme as n: where (7) RTT = RTT( N j and N are the set of actve ACKers whose new RTTs are receved n predefned calculatng tme and the number of, respectvely.. Acker Electon and Trackng (A) In order to track the hgh capacty recevers to be ACKers, we use normalzed throughput to track ACKers. The operaton of ACKers electon s performed by choosng the N recevers whose normalzed throughputs are hghest. We use the N hgh capacty recevers to be ACKers because the recevers could send ACKs quckly and effcently. Ths leads to the hgh achevable throughput. Ths selecton dffers from WET, whch selects the worse recever to be the ACKer. Ths s because n wred network, the sender adjusts the rate to the slowest recevers. The normalzed throughput derved n [6] and smplfed by [2] s calculated as: T nm ( = RTT( p( where RTT ( and p( are estmated round trp tme n (6) and estmated packet loss rate for recever j n (2), respectvely. To reduce the number of ACKer swtches, the selecton of new ACKers s shown as follow. or every endng of a predefned calculatng tme (n smulaton, we set the predefned calculatng tme to be 4 RTTs), the sender sorts the normalzed throughputs all actve recevers except the current ACKers. The sender consders only the frst N hghest throughput recevers (say, T ( x ),..., T ( x ) ) where x..., x N are the recevers n the frst N hghest throughput recevers. Also the sender sorts the normalzed throughputs of N current ACKers who pass a report test (say T ( y ),..., T ( y ), where y..., y N nm Anm nm N Anm (8) N are the ACKers ordered n descendng throughputs. The varable N s the number of ACKers passed the report test. The report test s performed by countng number of receved ACKs ( C ACK ( ) from ACKer, j durng the predefned calculatng tme, If C ACK ( s less than 50 percent of number of packet sent durng the predefned calculaton tme, the ACKer j does not pass the report test. Based on ths report test the predefned calculatng tme should be more than 2 RTTs. Ths report test s utlzed to solve the problem of ACKer falure. Then the sender elects N

6 CHUMCHU et al.: A MODEL-BAED CALABLE RELIABLE MULTICAT TRANPORT PROTOCOL 29 new ACKers by comparng the normalzed throughputs of the N recevers and the N current ACKers as: y y, x = TAnm ( y N, x, =,..., N ) ct nm, N = N ( x ), = N otherwse, = N N +,..., N +,..., N where c s constant between 0 and. If the value c s 0, the ACKers are fxed at the start connecton of the multcast sesson. On the other hand, f the value of c s, the ACKers are frequently changed. The protocol operaton for ACKer trackng s descrbed as follows. The sender sends 2 lsts: one for lst of dedcated ACKers and one for lst of current ACKers. Ths lst could be sent n form of a new packet or added to data packets. The frst lst s to notfy new ACKers and the last lst s to notfy current ACKers. If a recever receves the lst of dedcated ACKers consstng of ts address, t starts sendng ACKs. If the sender receves the frst ACK from the recever, t changes the lst of current ACKers by replacng old ACKer wth the new ACKer. If any current ACKers receve the lst of current ACKer not consstng of thers addresses, the ACKers stop sendng ACK. Therefore based on ths operaton, for any tme, there are at least N recevers who send ACKs. atellte Access de atellte Moble Earth taton AN() (9) VI. EXTENION OR MULTIPLE AN In ths secton, we provde the extenson for M atellte Access des shown n g. 6. Ths scenaro s deployed for satellte networks such BGAN [] and IPD [2]. We assume the sender knows servng AN of each recever. The congeston control s smlar to one AN scenaro. The source performs M wreless loss flter modules and M ACKer electons (A). One WLA and one A are nstalled for one AN. If the transmsson module receves a C (loss) from any WLA module, t sets up congeston wndow followng (4). The error control part performs the same as n one AN scenaro. VII. IMULATION REULT AND NUMERICAL REULT In ths secton, we present smulaton results of RMT over BGAN scenaro 2 usng N2 [3]. The detal of BGAN smulaton scenaro could be found n [9, 0]. or ths scenaro, we selected a bearer type 34 (422.4 kbps) for forward bearer. We fxed return bearer types to 76(68 kbps) for all smulated recevers. We used return slot plans, 7 and 9, whch also use 5-mllsecond slot szes []. Both common wreless losses and last mle wreless losses are modeled as temporally uncorrelated losses. mulaton results for temporally correlated losses could be found n [9]. The results of smulaton n ectons A, B, E,, and G are averaged from about 000 teratons. The results of smulaton n ectons C and D are collected from only one teraton. The objectve of ths smulaton s to present performance of RMT n terms of scalablty, achevable throughput and Goodput, the proposed wreless loss flter algorthm, and the proposed congeston control algorthm. ource 2 AN(2) Congest ons N atellte AN(N) g. 6 Multple-AN Network tructure g. 8 Probablty of congeston wndow unnecessarly reduced because of spatally uncorrelated wreless losses, p U A. calablty g. 7 Average Wreless Lnk Utlzaton per Orgnal Packet In ths subsecton, we present the scalablty of RMT protocol n terms of average wreless lnk utlzaton. g. 7 shows the average wreless lnk utlzaton per orgnal packet. Curve, curve 3 and curve 5 present RMT s smulaton results of average wreless lnk utlzaton per

7 30 JOURNAL O COMMUNICATION OTWARE AND YTEM, VOL., NO., EPTEMBER 2005 orgnal packet when last mle BERs of 0, 0-6 and 0-5, respectvely. Curve 2, curve 4 and curve 6 present RMT s numercal results detaled n [9] of average wreless lnk utlzaton per orgnal packet when last mle BERs of 0, 0-6 and 0-5, respectvely. Curve 7, curve 8 and curve 9 present N3 (wth K=) s numercal results of average wreless lnk utlzaton per orgnal packet when last mle BERs of 0, 0-6 and 0-5, respectvely. Curve 0, curve 2 and curve 4 present MTP s smulaton results of average wreless lnk utlzaton per orgnal packet when last mle BERs of 0, 0-6 and 0-5, respectvely. Curve, curve 3 and curve 5 present MTP s numercal results of average wreless lnk utlzaton per orgnal packet when last mle BERs of 0, 0-6 and 0-5, respectvely. Detals of MTP analyss could be found n [9]. It could be seen that the average wreless lnk utlzaton of RMT s vrtually ndependent of number of recevers and less than that of N3 (wth K=) and MTP. g. 9 RMT Goodputs versus numbers of ACKers for dfferent wreless loss rates n last mle lnk B. Congeston Control Part Ths experment s to show the mprovement of the proposed WLA and BWE for congeston control. g. 8 shows the probablty of congeston wndow unnecessarly reduced because of common wreless losses, p. Curve and U Curve 2 present the probablty of WET obtaned form analyss n [9] and smulaton. Curve 3 and Curve 4 shows the probablty of WLA obtaned form analyss n [9] and smulaton when number of ACKers s 2. Curve 5 and Curve 6 show the probablty of WLA obtaned form analyss n [9] and smulaton when number of ACKers s 3. Curve 6 and Curve 7 show the probablty of WLA obtaned form analyss n [9] and smulaton when number of ACKers s 4. The dfference between smulaton results and numercal results s that because n analyss, we do not consder packet losses durng tme out. It could be seen from the fgure, the probablty of WLA s much less than that of WET. or WLA, f the hgher the number of ACKers s, the less the probablty s. However, f the number of ACKers s too hgh, the feed back mploson occurs. Ths leads to the acheve Goodput s sgnfcantly reduced. Next experment s to nvestgate the effect of number of ACKers and congeston control wth bandwdth estmaton. The parameters of experment used are descrbed followng. The number of recevers s 0. The number of transferred packet s 000 packets. Packet sze s 500 bytes. mulaton results are shown n g. 9 and g. 0. Curve, curve 3, curve 5, and curve 7 show Goodputs of RMT wthout bandwdth estmaton when last mle BERs of 0, 0-6, 5x0-6 and 0-5, respectvely. Curve 2, curve 4, curve 6 and curve 8 show Goodputs of RMT wth bandwdth estmaton when last mle BERs of 0, 0-6, 5x0-6 and 0-5, respectvely. rom g. 9, t could be seen that the Goodputs of RMT wth BWE are hgher than those of RMT wthout BWE. Also, t could be seen that the number of ACKers of about 3, RMT gves the hghest Goodput for all packet loss rates. Ths s because f the number of ACKers s lower, the probablty p s hgher. U If the number of ACKers s too hgh, the feed back mploson occurs. g. 0 RMT Goodputs aganst number of ACKers for dfferent common wreless loss rates g. Intra protocol arness wthout wreless loss

8 CHUMCHU et al.: A MODEL-BAED CALABLE RELIABLE MULTICAT TRANPORT PROTOCOL 3 g. 0 shows RMT Goodput aganst number of ACKers for dfferent common wreless loss rates. Curve, curve 3, and curve 5 show Goodputs of RMT wthout BWE when frst mle BERs of 0, 5x0-6 and 0-5, respectvely. Curve 2, curve 4, and curve 6 show Goodputs of RMT wth BWE when frst mle BERs of 0, 5x0-6 and 0-5, respectvely. It could be seen that Goodputs of RMT wthout BWE are much less than those of RMT wth BWE. Ths s because n the case of common wreless loss, the WLA could not flter wreless losses out. Therefore to mprove Goodputs, RMT wth BWE performs much better. 32 packets. In order to verfy the behavor of competng TCP and RMT flows, we have run a large number of smulatons wth two types of flows. g. 4, g. 5 and g. 6 show the nstant throughput of TCP and RMT wthout wreless loss and wth wreless loss, respectvely. The legends of g. 4, g. 5 and g. 6 are : TCP nstance throughput wth maxmum wndow of 32 packets, 2: RMT nstance throughput wth maxmum wndow of 32 packets, 3: TCP nstance throughput wth maxmum wndow of 64 packets, 4: RMT nstance throughput wth maxmum wndow of 64 packets. C. Intra Protocol arness One of the requrement of relable multcast transport protocols [5], t should be far to other flows. In ths experment, we test 2 flows of RMT to test ts farness. The smulaton scenaro s the same as n g. The number of MEs s 6. The frst 3 MEs are attached RMT flow recevers. The lnk queue sze s 32 packets. The last 3 MEs are attached RMT flow 2 recevers. The number of ACKers s 3. The multcast sources of flow and flow 2 started at 2 second and 0 second, respectvely. g., g. 2 and g. 3 show the evaluaton of two competng RMT sessons through the same bottleneck wthout wreless loss, wth common wreless loss and wth spatally uncorrelated wreless loss, respectvely. Curve and curve 3 present nstance throughput and aggregated throughputs of flow. Curve 2 and curve 4 present nstance throughput and aggregated throughputs of flow 2. It could be seen that there s a good share of bandwdth. g. 3 Intra protocol arness wth spatally and temporally uncorrelated wreless loss n last mle lnk, BER=0-6 g. 2 Intra protocol arness wth common wreless loss BER=0-6 g. 4 Inter protocol arness wth non wreless loss D. Inter Protocol arness To show nter protocol farness, n ths subsecton, we test RMT performance n presence of competng a TCP flow. The smulaton scenaro s presented n g.. The number of MEs s four. The frst 3 MEs are attached RMT recevers. The last ME s attached TCP snk. The sources of RMT and TCP are attached at the source node. The lnk queue sze s E. ACKer Trackng In the case of one AN, f any current ACKers provde low performance such as hgh loss rate or connecton falure. The sender should elect new ACKers to report recever packets nstead of the problem ACKers. Ths experment tests the performance of ACKer trackng algorthm. The number of recevers s 0. The ntalzed number of Ackers s 3. The

9 32 JOURNAL O COMMUNICATION OTWARE AND YTEM, VOL., NO., EPTEMBER 2005 last mle wreless loss rate s 5x0-6. g. 7 shows the nstant throughput of the RMT wth dfferent c values. g. 8 shows average number of ACKers wth dfferent c values. It could be seen that the hgher the c value s, the hgher the average number of ACKers s. The hgher number of Ackers s, the lower the achevable throughput s. rom g. 8, the number of ACKers s hgh that ntalzed ACKers, because rght after ACKer swtchng, the old ACKer has to send ACKs untl t receves a release notfcaton from the sender. The sender sends the release notfcaton after recevng any ACK from the new ACKer. The processng of changng ACKers takes at least one round trp tme. g. 7 Throughputs for dfferent value of c, : c=0, 2: c=0.5, 3: c=.0 3 g. 5 Inter protocol arness wth common wreless losses wth BER=0-6 2 g. 8 Average number of ACKers for dfferent values of c, : c=0, 2: c=0.5, 3: c= g. 6 Inter protocol arness wth spatally and temporally uncorrelated losses wth BER= g. 9 RMT nstance throughputs and TCP nstance throughputs

10 CHUMCHU et al.: A MODEL-BAED CALABLE RELIABLE MULTICAT TRANPORT PROTOCOL 33. RMT throughputs compared to TCP throughputs In ths secton, we present RMT throughputs for dfferent wreless loss models. We compare the throughput of RMT to that of TCP. TCP s selected because t s the de facto standard for relable uncast transport n the Internet today. In smulaton of RMT, we tested 3 recevers. All recevers are ACKers. In smulaton of TCP, we tested one connecton. The common frst mle wreless BER s 0-6. g. 9 shows the comparson throughputs of RMT and TCP. Curve, curve 2 and curve 3 present nstance recevng throughputs of TCP when loss rates are 0-6, 5x0-6 and 0-5, respectvely. Curve 4, curve 5 and curve 6 present nstance recevng throughputs of RMT when loss rates are 0-6, 5x0-6 and 0-5, respectvely. It could be seen that our protocol performs much better than TCP. Ths s because our protocol reduces the probablty of congeston wndow unnecessarly reduced because of wreless losses. In addton, our protocol employs bandwdth estmaton to set congeston wndow and slow start threshold n a congeston epsode. g. 20 RMT Goodputs and MTP Goodputs when transmsson rate s less than forward bandwdth G. RMT Goodputs compared to MTP Goodputs In ths secton, we compare RMT Goodputs to MTP Goodputs. The transmsson rate n MTP s not dynamcally adapted to network condtons. It could set transmsson rate at the sender. The MTP transmsson rate s bps. The number of transferred packets s 5000 packets whose sze s 000 bytes. The results are shown n g. 20. Curve, curve 3 and curve 5 show the Goodputs of RMT when last mle loss BERs are 0, 0-6 and 0-5, respectvely. Curve 2, curve 4 and curve 6 show the Goodputs of MTP when last mle loss BERs are 0, 0-6 and 0-5, respectvely. It could be seen from the fgure that at BER=0, the Goodput of MTP s hgher than that of RMT because MTP could send full lnk all connecton tme whle RMT has slow start phase and congeston avodance phase. At BERs of 0-6 and 0-5, the Goodputs of RMT are sgnfcantly hgher than those of MTP. Ths s because RMT employs the proposed adaptve forward error correcton whle MTP s round (pass)-based transmsson wthout forward error correcton. VIII. CONCLUION AND UTURE WORK We have presented a model-based relable multcast transport protocol for satellte networks. The protocol s desgned to mprove relable multcast transport protocol performance n terms of scalablty, throughput and Goodput and also be far wth other traffc n the networks. Ths mprovement could be acheved by followng algorthms: calablty s mproved by usng the proposed adaptve forward error correcton. It has been shown n both smulaton results and numercal results that the average wreless lnk utlzaton and the average number of transmsson are vrtually ndependent to the number of recevers. eedback mploson s sgnfcantly reduced by usng a new hybrd adaptve forward error correcton and NAK-based AR n error control part and the proposed wndow-based congeston control usng a few hghest throughput recevers to send ACKs n congeston control part. mulaton results and numercal results show that the wreless lnk utlzaton s ndependent of the number of recevers. Drop-to-zero problem occurrng from last mle wreless losses s sgnfcantly reduced by usng the proposed wreless loss flter algorthm. The smulaton results and numercal results show that the probablty of congeston wndow unnecessarly decreased s much lower than that of TCP and WET. Common wreless loss s solved by usng settng slow start threshold and congeston wndow followng estmated bandwdth smlar to [8] when a loss occurs. The smulaton results show that the achevable throughput s sgnfcantly hgher that of TCP. Long delay s mproved by usng a new adaptve forward error correcton algorthm. Ths yelds that achevable RMT Goodput s much hgher than that of MTP. arness s acheved by mmckng TCP whch reles on packet-conservaton prncple. We are n process of desgnng multcast connecton management. In addton, we ntend to pursue the development of control theoretcal models that wll enable us to study the stablty of RMT as a functon of the varous systems parameters. REERENCE [] BGAN specfcaton, physcal layer INMARAT. [2] PD specfcaton, physcal layer INMARAT. [3] UCB/LBLN/VINT Network mulator (N), [4] G. Akkor, M. Hadjtheodosou, and. J. Baras. IP Multcast va atellte: A urvey Unversty of Maryland, College Park, Insttute for ystems Research Techncal Report TR_

11 34 JOURNAL O COMMUNICATION OTWARE AND YTEM, VOL., NO., EPTEMBER 2005 [5] Basu, P. and Kanchanasut, K., "A relable multcast protocol for undrectonal satellte lnk", In Proceedng of Applcatons and the Internet ymposum, Jan [6] Bhattacharyya,., Towsley, D., and Kuros, J., "vel loss ndcaton flterng approach for multcast congeston control," Computer Communcatons pecal Issue on Multcast Deployment, vol. 24, no. 5-6, 200. [7] Cao, G. and Wu, J., "Relable multcast va satelltes", In Proceedng of Internatonal Conference on Informaton Technology: Codng and Computng, Apr [8] Casett, C., Gerla, M., Mascolo,., anadd, M. Y., and Wang, R., "TCP Westwood: End-to-End Congeston Control or wred/wreless Networks," Wreless Networks, vol. 8, no. 5, [9] P. Chumchu and A. enevratne. Analyss of relable multcast protocols n satellte networks Inmarsat Report. [0] P. Chumchu and A. enevratne. N-based GEO satellte network smulator INMARAT Report. [] ang, J. and Akyldz,. I., "A relable multcast transport protocol for satellte IP networks", In Proceedng of IEEE GLOBECOM 2002, v [2] loyd,., Jacobson, V., Lu, G. C., McCanne,., and Zhang, L., "A relable multcast framework for lght-weght sessons and applcaton level framng," IEEE/ACM Transactons on Networkng, vol. 5, no. 6, 997. [3] Jacobson, V., "Congeston avodance and control", In Proceedng of ACM IGCOMM, Aug [4] Jung, M, nenmacher, J., and Bersack, E. W., "Relable multcast va satellte: un-drectonal vs. b-drectonal", In Proceedng of KV'99, Mar [5] A. Mankn, A. Romanow,. Bradner, and V. Paxson. IET Crtera for evaluatng relable multcast transport and applcaton protocols. RC [6] Matthew Maths, Jeffrey emke, Jamshd Mahdav, and Teuns Ott, "The Macroscopc Behavor of the TCP Congeston Avodance Algorthm," Computer Communcatons Revew, 997. [7] K. Mller and K. Robertson. tarburst Multcast le Transfer protocol (MTP) specfcaton IET, IET-Draft <draft-mller-mftp-spec03. txt>. [8] P.Chumchu, R.Borel, and A.enevratne, "A New calable Relable Multcast Transport Protocol for atellte Networks (RMT)", In Proceedng of oftcom 2005, ep [9] P.Chumchu, R.Borel, and A.enevratne, "Performance Analyss of relable multcast transport Protocol for GEO atellte networks", In Proceedng of MACOT 2005, ep [20] Paul,., abnan, K. K., Ln, J.-H. C., and Bhattacharyya,., "Relable multcast transport protocol (RMTP)," IEEE Journal on elected Areas n Communcatons, vol. 5, no. 3, Jan [2] Rzzo, L., "pgmcc: a TCP-rendly sngle-rate multcast congeston control scheme", In Proceedng of ACM IGCOMM, Jan [22] ano, T., hroshta, T., Takahash, O., and Yamashta, M., "Montorng-based flow control for relable multcast protocols and ts evaluaton", In Proceedng of Performance, Computng, and Communcatons Conference, 997. [23] Wang, A. H. and chwartz, M., "Achevng Bounded arness for Multcast and TCP traffc n the Internet", In Proceedng of ACM IGCOMM, Jan [24] Wang, C. and Leung, V. C. M., "A relable multcast transport protocol for broadband satellte networks", In Proceedng of IEEE 2002 Internatonal Conference on Communcatons, Crcuts and ystems and West no Expostons, Jun Prawt Chumchu s a lecturer at Mahanakorn Unversty of Technology. He receved hs Bachelor degree (B.Eng) wth honour from Thammasat Unversty and Master degree (M.Eng) from Mahanakorn Unversty of technology, Thaland. He has a Ph.D. from Unversty of New outh Wales, Australa. Hs research nterests are n dgtal sgnal processng for telecommuncaton, telecommuncaton systems and moble computng. Dr Roksana Borel s a enor Research Engneer n NICTA s Networks and Pervasve Computng (NPC) and an Assocate Adjunct Professor at the chool of Electrcal and Telecommuncaton Engneerng, Unversty of New outh Wales (UNW). Aruna enevratne s the program leader of the networks and pervasve computng program at the Natonal ICT Australa. He holds the Mahanakorn Char of Telecommuncaton n the chool of Electrcal Engneerng and Telecommuncaton at the Unversty of New outh Wales. He has a Ph.D. from the Unversty of Bath, UK, and a B.c (Hons) degree from Mddlesex Polytechnc. nce graduatng he has held academc appontments at the Unversty Bradford (UK), Curtn Unversty, Australan Defense orce Academy (UNW) and the Unversty of Technology, ydney. Outsde academa he as worked at the tandard Telecommuncaton Laboratores (UK), Murhead and Co (UK), and Telecom Australa (Telstra). He has also spent tme at the MAI Laboratory at the Unversty of Perre Mare Cure (Pars), and the Rodeo Group at INRIA (Nce) as a vstng professor. Hs current research nterests are n ualty of ervce Management partcularly n moble data communcaton systems.