Rerouting Time and Queueing in Proactive Ad Hoc Networks

Transcription

1 Rerouing Time and Queueing in Proacive Ad Hoc Neworks Vinh Pham 1, Erlend Larsen 1, Knu Øvshus 2, Paal Engelsad 1 and Øivind Kure 3 1 UniK, Norway 2 Bergen Universiy College, Norway 3 Q2S, NTNU, Norway 1,3 {vph, erl, paalee, okure}@unik.no 2 knu.ovshus@hib.no Absrac In a MANET nework where nodes move frequenly, he probabiliy of conneciviy loss beween nodes migh be high, and communicaion sessions may easily loose conneciviy during ransmission. The rouing proocol is designed o find alernaive pahs in hese siuaions. This rerouing akes ime, and he laency is referred o as he rerouing ime. This paper invesigaes he rerouing ime of proacive rouing proocols and shows ha he rerouing ime is considerably affeced by queueing. Simulaions and analysis are conduced o explore he problem. Finally, we propose a MAC-layer soluion ha reduces he rerouing ime problems due o queueing. Simulaions and analysis show ha he soluion is so effecive ha i eliminaes he enire problem in many siuaions. 1. Inroducion The research effors in he field of ad hoc neworking have been going on for many decades. Ad hoc neworking enables communicaion direcly beween nodes, wihou he need for exra infrasrucure. This makes i very suiable for miliary and rescue operaions. The sandardizaion of rouing proocols has been underaken by he Mobile Ad Hoc Neworking (MANET) working group in IETF [1]. They are se o bring forward wo proocols, one reacive and one proacive. A common characerisic of ad hoc neworks is ha links may break due o changes in radio condiions, node mobiliy and oher ypes of nework dynamics. The rouing proocol is designed o find alernaive pahs in hese siuaions. The ime period before new pahs are found is referred o as he rerouing inerval, and he duraion of he rerouing inerval is referred o as he rerouing ime. During he rerouing inerval, sale roues exis over he link ha has been broken. Rerouing can only ake place afer he rouing proocol has deeced ha he link is broken. In fac, a significan par of he rerouing ime is associaed wih he deecion of he link break. Wih proacive rouing proocols, such as Opimized Link Sae Rouing (OLSR) and Open Shores Pah Firs wih MANET Designaed Rouers (OSPF-MDR), a link is mainained by he exchange of conrol packes. A link break is normally no deeced unil eiher a cerain number of HELLO packes have been los, or he lack of periodic updaes resuls in a link imeou [2-4]. (Some implemenaions migh le he link layer deec link breaks and signal his informaion o he rouing proocol. Such cross-layer opimizaions are ouside scope of his paper. Here, we explore he common layered approach where HELLO packes are necessary for he deecion of link breaks.) Wih he defaul parameer seings of OLSR and OSPF-MDR, a link break should normally be deeced afer approximaely 6 seconds. However, we conduced a series of lab experimens of OLSR [3] and OSPF- MDR [4] and observed rerouing imes ypically in he order of seconds. Since he rerouing ime depended on ransmission raes of daa raffic and on size of he ransmission queues, we realized ha he increased rerouing ime in our experimen was mainly caused by he queueing of he daa packes. During he rerouing inerval, he nework layer a he node upsream o he broken link migh ry o forward daa packes over he broken link. Insead, hese packes are accumulaed in he oupu queue. Due o he layered design, he link layer (L2) will keep rying o ransmi he queued daa raffic already designaed o he broken link, even afer he nework layer (L3) has imed ou he link. This does no only consume scarce radio resources. I also blocks he MAC layer. Thus he nework layer is no able o announce ha he link is broken, and he rerouing ime increases correspondingly. Finally, when all he sale daa packes designaed o he oupu queue have been dropped, he MAC layer is ready o ransmi he link sae announcemen o esablish new roues hroughou he nework and o

2 hp://folk.uio.no/paale serve packes waiing in he oupu queue designaed o reachable receivers. In summary, he rerouing ime due o link breaks depends on he ime o carry ou he following processes: Deecion of a link break The empying of all sale packes from he oupu queue Nework-wide link-sae announcemen o esablish new pahs While boh link break deecion and rouing convergence have received considerable aenion in he research communiy, surprisingly lile focus has been direced o he effecs of queueing. Indeed, he main conribuion of his paper is o explore how queueing increases he rerouing ime. The res of he paper is organized as follows. Secion 2 gives background informaion on relevan echnologies. In Secion 3 we presen he simulaion seup, define he rerouing ime, and show simulaion resuls. Secion 4 gives an analysis of he facors conribuing o he rerouing ime. Secion 5 presens a proposed soluion o he rerouing problem and in Secion 6 we presen some relaed work. Finally, in Secion 7 he conclusion is presened and furher work is skeched ou. 2. Background 2.1. The MAC layer of IEEE Today, IEEE [5] is he mos widely used wireless local area neworking echnology. The sandard defines a Physical (PHY) layer and a Medium Access Conrol (MAC) sub-layer, where he laer suppors wo modes of operaion, namely he Disribued Coordinaion Funcion (DCF) and he Poin Coordinaion Funcion (PCF). Since DCF is he mos common mode of operaion, we focus only on DCF in his paper. Wih DCF, he wireless saions (STAs) access he medium in a disribued way, using carrier sense muliple access wih collision avoidance (CSMA/CA). Wih he basic access mechanism, each unicas DATA frame is acknowledged wih an ACK frame. This is also known as he minimal frame exchange. (Mulicas ransmissions, however, are no followed by an ACK frame.) Wih he opional 4-way frame exchange, on he conrary, each DATA frame is preceded by an exchange of a reques o send (RTS) and a clear o send (CTS) frame. The use of RTS/CTS is paricularly useful o avoid collisions due o hidden erminals [6]. Source Desinaion Oher Fig. 1. Basic CSMA/CA. The basic access mechanism wih he minimal frame exchange is illusraed in Fig. 1. When a node has some daa o ransmi, i has o sense he medium o verify wheher i is busy or idle. If he channel is idle for a ime inerval equal o a Disribued Iner-Frame Space (DIFS), he source node may begin daa ransmission. When he receiver has received he DATA frame, i wais for a ime inerval equal o a Shor Iner-Frame Space (SIFS), and ransmis an ACK back o he source node. While daa is being ransmied, all oher nodes mus defer heir channel access for a ime inerval equal o he Nework Allocaion Vecor (NAV). This is a imer indicaing he amoun of ime ha he medium has been reserved for he curren ransmission. When he daa ransmission is finished and he NAV has expired, a new conenion period is enered. Here, concurren nodes wih pending daa raffic mus conend for he medium. In his process, each conending node mus choose a random ime inerval called Backoff_imer, seleced from he conenion window (CW) in he following way: Backoff _ imer= rand, where DIFS DATA CW = [ CW min, CWmax]. NAV SIFS Defer access ACK DIFS [ 0 CW] sloime CW Backoff afer defer The value for he sloime is dependen on he PHY layer ype. The backoff imer is decremened only afer each ime he medium is idle for a DIFS inerval, and is frozen when he medium becomes busy. Evenually when he backoff imer of a node expires, i migh ransmi daa. The main poin of his medium access mechanism is o minimize he probabiliy of a collision, i.e. of concurren ransmissions. Since a node mus go hrough a backoff afer having ransmied a frame (also referred o as a pos-backoff), he medium access hp://

3 mechanism also provides long erm fairness o access he medium. In a wireless environmen where collision deecion is hard or even impossible, a posiive ACK from he receiver is used o confirm a successful ransmission. The absence of such an ACK message indicaes a collision, link failure or oher reasons for an unsuccessful ransmission. When his occurs, a reransmission is scheduled, and a new backoff value is chosen. However, in order o reduce he risk for consecuive collisions, afer each unsuccessful ransmission aemp, he CW is doubled unil a predefined CW max is reached. There is a rery couner associaed wih he ransmission of each frame, and he rery couner is incremened afer each collision. Afer a successful reransmission, he CW is again rese o a predefined CW min, and he rery couner is rese o null. The maximum number of reransmissions for a frame is defined in he do11shorrerylimi and do11longrerylimi variables. The firs variable is applicable for MAC frames ransmied wih he minimal frame exchange (i.e. wih lengh less han or equal o he do11rtsthreshold parameer), while he laer is applicable o frames ransmied wih RTS/CTS. For insance, each ime a MAC frame of lengh less han or equal o he do11rtsthreshold is ransmied, and i fails, he shor rery couner is incremened. This will coninue unil here is a successful ransmission or he couner has reached he do11shorrerylimi and he packe is discarded. When his happens he shor rery couner is rese o zero. For simpliciy, hroughou he res of his paper, we will use he erm do11shorrerylimi and rery limi inerchangeably Queueing in he proocol sack The unicas packes (mulicas is considered ou of scope) creaed by applicaions are passed down he proocol sack o TCP or UDP using he socke inerface (Fig. 2). If he packe is a TCP packe, i may be queued o accommodae flow conrol. For UDP, and TCP evenually, he packe is passed down o L3 (i.e. he IP layer) for rouing and designaion of a nex hop link layer address before passed down o he L2 (i.e. he link layer). There i is queued in he queue of he device driver unil he buffer of he nework inerface is empy, and is hen pulled ono he nework inerface. When he ransmission medium is available, he packe is ransmied. If no ACK is received, he packe is assumed los due o a collision, and he packe will be scheduled for reransmission. Socke/ INET TCP/ UDP IP Device Driver Applicaion User Space Kernel Space Hardware Applicaion Fig. 2. Linux proocol sack [7]. When a packe is received a an inerface, i is pu in a backlog queue. Then L3 processes i, and eiher forwards i ou on an inerface or pushes i up he sack o UDP or TCP. TCP has a receive queue o serve flow conrol. The L2 queue should be of a minimum size o allow raffic o be sen wihou loss from applicaions a a rae higher han he nework capaciy, as he nework bandwidh can be variable due o fading, mobiliy, inerference, conenion ec. Boh Linux and he nework simulaor ns-2 [8] implemen a L2 queue for ougoing packes. In ns-2, using he CMU Monarchs wireless exensions, packes are queued in he inerface prioriy queue (IFq). The nework sack for a mobile node consiss of a link layer (LL), an ARP module conneced o LL, an inerface prioriy queue, a MAC layer, a nework inerface, all conneced o he channel. When a packe is creaed by he source applicaion, he packe is queued in he IFq unil all previous packes have been eiher sen or discarded Opimized Link Sae Rouing Proocol (OLSR) OLSR is a proacive rouing proocol for ad hoc neworks. The proocol is buil around he noion of Muli Poin Relay nodes (MPRs). The main purpose of MPRs is o creae and forward link sae messages. The MPRs are seleced individually by each node in he

4 nework in such a way ha all nodes can reach heir 2- hop neighbor nodes hrough an MPR. The wo mos imporan message ypes in OLSR are he HELLO and he TC (Topology Conrol) messages: 1) HELLO Messages: Every node broadcass HELLO messages periodically, o suppor link sensing, deecion of neighbors and signaling of MPR selecion. The recommended emission inerval for HELLO messages is 2 seconds, and he holding ime for neighbor informaion is 6 seconds. Thus a neighbor is considered los 6 seconds afer he las HELLO message received from he neighbor. 2) TC Messages: Based on he informaion colleced hrough HELLO messages, link sae (TC) messages are creaed and broadcased hroughou he nework by each MPR. The recommended emission inerval for TC messages is 5 seconds, and he holding ime is 15 seconds. 3. Simulaions 3.1. Descripion of he scenario The scenario explored can be described as follows: Three nodes A, B and C form an ad hoc nework where A sends raffic o C a a Consan Bi Rae (CBR). A he beginning, A and B srech ou he nework. Then C moves pas B, and loses conneciviy wih A unil raffic from A is reroued via B. In his scenario, C has always direc conneciviy wih B. Alhough he scenario seems simple, i is realisic and sufficien o explore imporan aspecs of he rerouing ime. Noe also ha all nodes are wihin a wo-hop disance of each oher. This means ha he disseminaion of TC messages will no affec he rerouing ime, and we are able o explore he rerouing ime associaed only wih he deecion of he link break and wih he queueing effecs. Fig. 4. Concepual model for he simulaions Simulaion seup All simulaions were carried ou using he nework simulaor ns-2. The seup for our es scenario (Fig. 4) is equivalen o he scenario in Fig. 3. In he beginning, all hree nodes A, B and C are in he immediae neighborhood of each oher. Node A, which is he sender, sends UDP daa packes of rae R in packes per second direcly o he receiver node C. (This flow is marked wih "(1)" in Fig. 4.). While he daa ransmission is ongoing, node C moves away from node A. A a cerain poin where node A and C are no longer in he immediae neighborhood of each oher, he connecion beween hese wo nodes is broken. In order o re-esablish conneciviy beween node A and node C, node A has o reroue he raffic hrough node B. (This flow is marked wih "(2)" in Fig. 4.). B forwards his raffic furher on o node C. In all simulaions, he packe size was fixed a 1000 byes. IEEE b [9] was used wih he basic DCF mechanism (i.e. wihou RTS/CTS) and a nominal ransmission rae of 11 Mbps. The RTS/CTS handshake mechanism is no necessary since here is no hidden node problem in our scenario. All nodes are inside each ohers sensing range. Table 1. Simulaion parameer seings. Simulaor ns-2 version 2.30 Radio-propagaion model TwoRayGround MAC ype b Inerface queue ype FIFO wih DropTail Anenna model OmniAnenna Daa rae 11 Mbps Basic rae 1 Mbps Packe Size IP 1000 Byes Movemen speed of node C 3.3 m/s OLSR HELLO_INTERVAL 2 seconds OLSR REFRESH_INTERVAL 2 seconds OLSR TC_INTERVAL 5 seconds OLSR NEIGHB_HOLD_TIME 6 seconds OLSR TOP_HOLD_TIME 15 seconds OLSR DUP_HOLD_TIME 30 seconds The implemenaion of OLSR by he Universiy of Murcia was used as he proacive rouing proocol for ns-2 [10]. In he OLSR configuraion, he ime inerval beween HELLO packes was se o 2 seconds, and he HELLO packes were given prioriy over daa packes o avoid roue insabiliy. Furhermore, a link is considered down afer he loss of 3 consecuive HELLO packes, leading o a deecion ime of link breaks of approximaely 6 seconds: HELLO_ INTERVAL= 2seconds NEIGHB_ HOLD_ TIME= 3 HELLO_ INTERVAL Essenial parameers used in he simulaions seup are summarized in Table 1.

5 3.3. Definiion of he rerouing ime In he simulaions ha were conduced, we mainly focused on measuring he rerouing ime, i.e. he ime duraion from when he link beween A and C is broken o he ime when conneciviy is re-esablished via he inermediae node B. However, our experience hrough many experimens - boh in a real es-bed and in simulaions - is ha he rerouing ime measured in his way will have a high degree of variance caused by random effecs during rerouing. In order o minimize variance in he measuremens, we have chosen o define he rerouing ime reroue as he ime inerval from he las HELLO message from node C received by node A before link break, o he momen where he conneciviy is re-esablished, i.e. unil he insan of ime where he firs UDP packe is received a C afer he link break. (marked as crosses, squares and riangles). Here, he rery limi is se o 7, and he queue size is se o 100, 400 and 1000 packes. The figure shows ha for small packe raes, he rerouing ime is a he minimum value of 6 seconds, which equals o he NEIGHB_HOLD_TIME. As he packe rae increases, he rerouing ime also increases linearly up o a cerain poin where i suddenly sops o increase, and he rerouing ime sabilizes a is maximum value. The maximum rerouing ime depends on he queue size. For a queue size of 100, he maximum rerouing ime is slighly more han 10 seconds. For a queue size of 400 i is nearly 23 seconds, while for a queue size of 1000 he maximum rerouing ime lies around 47 seconds. Wih a queue size of 400, we see ha a packe raes of 100 pks/sec and over, he queue is filled a he ime when rerouing akes place, and his resuls in a rerouing ime converging on approximaely 23 seconds Simulaion resuls The resuls from he simulaions for various rery limis (Fig. 5) show ha a higher rery value gives a longer rerouing ime. This is as expeced, because each packe in he L2 queue is ransmied a number of imes defined by his rery value. We also noice ha he rerouing ime is linearly proporional wih he L2 queue size. Rerouing ime (sec) Theoreical Ifq 100 Theoreical Ifq 400 Theoreical Ifq 1000 Simulaion Ifq 100 Simulaion Ifq 400 Simulaion Ifq 1000 Rerouing ime (sec) do11shorrerylimi 1 do11shorrerylimi 4 do11shorrerylimi 7 do11shorrerylimi Packe Rae (pks/sec) Fig. 6. Simulaion resuls for rerouing ime over packe rae for layer 2 queue size 100, 400 and 1000 packes, 7 MAC reries. (95% conf. in.) I is also observed (Fig. 6) ha a low raes (i.e. well below 20 pks/sec) he rerouing ime is fla a 6 seconds Layer 2 queue size (packes) Fig. 5. Simulaion resuls of rerouing ime over layer 2 queue size. Fig. 6 shows simulaion resuls for he rerouing ime as a funcion of he ransmied packe rae 4. Analysis 4.1. Analysis of he problem From he log file produced by ns-2 we can observe various incidens affecing he rerouing ime. These incidens, which occur a node A, are illusraed in Fig. 7 and are explained below:

6 break beween A and C, A sill has a roue o C in is 2- hop neighbor able, i.e. hrough node B. Therefore, node A does no need o wai for any TC message from B in order o figure ou how o reach C. Fig. 7. Illusraing differen incidens a node A s queue as funcion of ime. 1) In region I, daa packes are coninuously insered ino he ransmi queue a node A. A ime 0, he las HELLO message from C is received a A (shor line in he figure). Then, afer b seconds, he direc link beween A and C is broken a 1. 2) Alhough he link beween A and C is broken, he rouing proocol is sill no aware of his, and herefore has no updaed he rouing able. As a resul, garbage daa packes wih sale rouing informaion coninue o be pu ino he queue a node A. 3) In region II, i.e. 1 < < 2, he queue a node A is being filled up. This happens since each garbage daa packe in he queue a node A is reransmied L imes, where L is he rery limi. Because of all he reransmissions for each packe, he packe rae ou of queue R ou will be reduced considerably. As long as he packe rae R in ino he queue is higher han R ou, he queue will be filled up. This will las for d b seconds, where d is he imeou value for he rouing proocol s HELLO packes (which is equivalen wih NEIGHB_HOLD_TIME in OLSR). 4) A 2, garbage packes are no longer pu ino he queue. The rouing proocol has now updaed he rouing ables. New daa packes are insead correcly reroued o B. 5) In region III, i.e. 2 < < 3, he queue is being empied for garbage packes. This will las for e seconds, depending on parameer values like R in, L, packe size, queue size ec. Noe ha packes are aemped ransmied and removed from he queue boh in region II and region III. Thus, he queue will fill up in region II only if R in >R ou. However, for he lowes packe raes, we will have R in <R ou, and he queue will no be filled. In he laer case, boh region II and region III will be nonexisen, and he rouing inerval consiss of only region I. This explains why he rerouing ime is fla a 6 seconds for he lowes packe raes in Fig. 6. In summary, he incidens in he ime inerval 0 < < 3 are he main conribuions o he rerouing ime as defined above. I is also worh noing ha in our es scenario, he delay from a TC message is irrelevan for he rerouing ime. This is due o he fac ha prior o he link break, node A will have node C in boh is 1- hop and 2-hop neighbor ses. When A discovers a link 4.2. A model for he rerouing ime Based on he observaions above we have derived a simple analyical model ha can be used o predic he rerouing ime. The derivaion of he model is given below: difs DCF inerframe space bo backoff ime daa delay for ransmiing he daa packe sifs shor inerframe space ack delay of acknowledge T e slo ime in IEEE RTS delay of a RTS packe CTS delay of a CTS packe L number of reries B queue size 1) According o he sandard, he delay of aemping o ransmi a single packe over a broken link is T + packe = (1) c bo where T c is he delay associaed wih he ransmission aemp, and bo is he delay associaed wih he backoff. In our scenario, no ACK is received when he link is broken, and he ransmission aemp is herefore perceived as a collision. However, according o he sandard, a node mus wai an ACKTimeou amoun of ime wihou receiving an ACK frame before concluding ha he ransmission failed. In our case, his ACKTimeou corresponds o he ransmission of an ACK for a successfully ransmied frame. Thus, he delay associaed wih he ransmission aemp, T c, is equal o he delay associaed wih a successful ransmission, T s : T T = c =. (2a) s difs daa sifs ack Wih he RTS/CTS mechanism, on he conrary: T c = + +. (2b) difs RTS CTS _ imou Prior o each packe ransmission, a backoff ime is uniformly chosen in he range (0, W j -1). Here we

7 define W j, where j ( 0, m), as he conenion window a backoff sage j, and m is he number of he maximum backoff sage. Le us also define L as he number of reries, and we can hus wrie he definiion of he conenion window as: j 2 W0 W j = m 2 W0 L m L> m. (3) Eq. (3) saes ha for he firs ransmission aemp, he conenion window is W 0 which is equal o CW min. (Noe ha his definiion of he conenion window is slighly differen from he definiion in Secion 2. In fac, he IEEE sandard refers o W j -1 as he conenion window [5]. For convenience, we have defined he conenion window differenly in his paper.) Afer each unsuccessful ransmission, he conenion window is doubled, and he packe is aemp reransmied. This will coninue unil we reach he maximum conenion window W m = 2 m W 0 = CW max, where i remains for consecuive reransmission aemps. If a reransmission is successful afer a number of reries, or he number of reransmission has reached he rery limi, he conenion window is again rese o is iniial backoff sage W 0. In our scenario, when he link beween A and C is broken, each garbage packe in he queue is reransmied L imes, and evenually is discarded because he maximum number of reries has reached. The mean oal delay for one single packe wih L reries is hen approximaely: + packe L = ( L+ 1) Tc bo where _ (4) L+ 1 [ 2 1] W0 1 ( ) ( L+ 1) L m (5) 2 2 bo = Te W0 m 1 ( ) [ 2 ( L m+ 2) 1] ( L+ 1) L> m 2 2 which is he sum of he approximae mean backoff ime. Here, T e is he slo ime. Noe ha T e, W 0 and m are parameers ha depend on he PHY-layer used. For b, T e = 20 µs, W 0 = 32 and m = 5. We have inenionally ried o keep he scenario as simple as possible, o derive a simplified model ha is inuiive and easy o analyze. One of he simplificaions made is he assumpion ha A is he only node rying o access he medium when he link is broken. Thus, during backoff he medium is always idle, and he duraion of each backoff sae is herefore T e. I is no difficul o exend our analysis for he case when muliple nodes conend for he same medium. In [11], for example, Engelsad and Øserbø calculaed he queueing delay by applying a Bianchi model ha is exended o non-sauraion condiions. Thus, exending our analysis is no hard o do, bu draws aenion away from he main objecive of his paper. I is also considered ou of scope due o space limiaions, bu migh be addressed in a follow-on publicaion. 2) The packe rae R ou ou of queue when each packe has o be reransmied L imes, is herefore: R ou = min 1 packe_ L, R in 1 = min ( L+ 1) Tc + 3) The oal rerouing ime is: rerouing d e bo, R in. (6) = +, (7) where (depiced in Fig. 7): 1 e = min[ ( d b ) ( R in R ou ), B]. (8) R ou 4.3. Discussion Eq. (7) equals he rerouing ime as defined above, where only he mos significan mechanisms conribuing o he oal delay of he rerouing ime is considered. This delay is equal o 3 0 in Fig. 7. Here, we assume ha he delay of ransmiing one single packe hrough he alernaive pah, from A o B and hen o C, is very small compared o d and e. This delay is herefore omied in he equaion. The firs erm of he equaion is a consan defined by he proacive ad hoc rouing proocol configuraion (his is equivalen o he NEIGHB_HOLD_TIME in OLSR). This value is also he absolue minimum rerouing ime. The second erm is variable, depending on parameers like R in, he rery limi L, he queue size B, ec. From Eq. (7) and Eq. (8) i is clear ha here is a lower and an upper limi on he rerouing ime. The lower limi occurs when R in = R ou, in Eq. (8). Thus, for he lowes packe raes he rerouing ime is equal o d, as we also observed in he simulaions. The upper limi occurs when he queue is filled and is consrained by he queue size B. Hence, for he highes packe raes (i.e. when R in > B/( d b )+R ou ) he maximum rerouing ime is:

8 B rerouing_max = d +. (9) R ou Furhermore, in he case when he rerouing ime is larger han d and smaller han rerouing_max, Eq. (7) yields: R in rerouing = ( d b ) + b. (10) R ou This reveals ha he rerouing ime is linear and proporional o R in in his region. Table 2. Comparison of he delay componens of R ou and he resuling value for R ou. Values are given in milliseconds for he delay erms. R ou is given as packes per second. L bo (L+1)T s R ou The packe rae ou of he ransmi queue R ou is also an imporan parameer for he rerouing ime. A decreasing R ou means an increasing rerouing ime. By inspecing Eq. (6), we see ha he firs erm in he denominaor is linearly proporional wih L, while he second erm is increasing exponenially wih L [Eq. (5)]. This means ha he second erm will grow much faser han he firs erm, and herefore will be he dominaing erm when L is large. This is illusraed in Table 2 where only he resuls for he eigh firs rery values were calculaed. Here, a packe size of 1000 byes wih a ransmission rae of 11 Mbps was used o calculae he delay of daa (in T c ) in Eq. (6). The rae in R in was se o 100 packes per second. The resuls from Table 2 show ha for he given seing, R ou is rapidly decreasing for rery values above 4. A plo of he esimaed rerouing imes based on Eq. (7) is shown for hree differen queue sizes (100, 400 and 1000 packes) as dashed curves in Fig. 6. The curves were calculaed using a value of b = 0.2 seconds, which corresponds o he b value observed in he simulaion resuls shown in he figure. As he resul shows, he esimaed rerouing imes are almos equal o he simulaed resuls obained from ns-2. This verifies ha he derived formula is a good approximaion for he expeced rerouing ime in he given scenario. We observe, however, ha he simulaions give a slighly higher rerouing ime. This can be explained by he ARP reques burs riggered by all packes sen o he Layer 2 in he ime lapse from he roue hrough B is chosen, unil node B s MAC address is obained. This behavior of ns-2 is a violaion of he recommendaions given in [12]. The ARP sorm problem is bigger for higher packe raes and larger queue sizes, which can be observed in he figure. 5. Proposed soluion 5.1. Adapive rery limi A he ime he rouing proocol becomes aware ha he direc connecion o he desinaion has been broken, he packes in he L2 queue no longer have a reachable link layer desinaion. These packes will be discarded only afer being ransmied ono he medium for a number of imes defined by he IEEE do11shorrerylimi. We argue ha a soluion o his problem should be implemened as a layered soluion, o keep i as small and simple as possible. The link layer proocol will be able o deec he link break earlier han he rouing proocol, so i is naural o implemen a soluion a he link layer. Our analysis shows ha a he link layer i is he queue size and he rery limi ha are he main conribuors o he exended rerouing ime. Reducing he queue size could be an opion, bu o have any effec, his reducion would have o be iniiaed as soon as he queue usage sars o grow. In his case i would be more efficien o keep he queue small a all imes, insead of varying i, bu his would resrain he flexibiliy of having a large queue. Insead, we propose a soluion o he accumulaed queue ime problem by inroducing an adapive rery limi ino he IEEE DCF MAC. For each successive packe wih he same desinaion MAC address ha is discarded due o reaching he rery limi, he rery limi is reduced by 1, unil each packe is only aemped ransmied 1 ime. If he original rery limi is 7, he rery limi is reduced o 0 afer 7 consecuive packes are dropped due o reaching he rery limi. As soon as a packe is ransmied successfully, he rery limi is rese o is original value equal o ha of he legacy IEEE sandard Discussion I is very rare ha many rery couner expiraions occur direcly following each oher, unless somehing is wrong. To lower he rery limi gradually will probably no affec he funcionaliy of he MAC under normal nework condiions. However, a problem wih he adapive rery limi soluion is ha i migh lead o

9 an unfair resource disribuion in erms of collision avoidance. The node sending packes ha go unacknowledged will be able o conend for he medium wih a high probabiliy of a smaller backoffcouner han is peers. On he oher hand, he empying of sale packes from he queue akes place in a small period of ime, and i is much more efficien o send a garbage packe only one ime, han sending i muliple imes. Anoher drawback is ha he ransmission aemps of garbage packes consume nework resources. More complex soluions where he MAC layer discards packes wihou aemping o ransmi hem is cerainly also possible. In summary, here are a number of variaions of he proposed adapive rery limi soluion. The performance of a number of hese variaions in various neworking scenarios will be deailed and discussed in a follow-on publicaion Implemenaion To do he acual implemenaion in ns-2 we needed o inroduce wo new variables. The firs of hese new variables, ID, keeps rack of he desinaion of he las ransmission aemp, and he second variable, called PCn, couns he number of packes discarded because he rery couner has reached he rery limi. Each ime a packe is discarded because he rery couner has reached he rery limi, he PCn is increased, unil i reaches he value of he rery limi. The PCn is subraced from he original rery limi, so ha he effecive rery limi ges lower and lower as he PCn increases, unil new packes are only ransmied once, and hen discarded if no acknowledged by he receiving node. If a packe is ransmied o a new desinaion, PCn is se o 0 and ID is updaed. If he ransmission was successful (indicaed by a received ACK), he PCn is se o Simulaion resuls In he simulaion resuls of he adapive rery limi soluion (Fig. 8, wih L2 queue size 400 packes and 7 MAC reries) we observe ha wih he proposed soluion he rerouing ime is kep a 6 seconds (which equals o NEIGHB_HOLD_TIME) unil he packe rae exceeds 600 pks/sec. A his packe rae he bi rae approaches he heoreical maximum hroughpu (TMT) of 5.03 Mbps (for 1000 byes sized packes, and for a nework wih one sender, where backoff ime has o be aken ino accoun). When he packe rae is higher han TMT, he L2 queue ges filled also when he link beween A and C is no broken. This is because R ou is smaller han R in a all imes. Rerouing ime (sec) Theoreical Ifq 400 Theoreical Adapive Ifq 400 Adapive Ifq 400 Simulaion Ifq Packe rae (pks/sec) TMT Fig. 8. Simulaed resuls for soluion wih adapive rery limi. (95% conf. in.) Our soluion prevens R ou from decreasing when he link beween A and C is broken, by limiing he accumulaed number of reransmissions. The resuls from he simulaions have proven ha he adapive rery limi soluion effecively eliminaes he delay relaed o he queueing problem. The proposed soluion is a layered approach based a he link layer, bu in some cases i would be more convenien o solve he problem a he IP-layer. This is lef for furher work. 6. Relaed work An analysis of several neighbor sensing approaches is presened in [13]. The objecive is o beer be able o opimize performance in an OLSR nework. In [14], he OLSR rouing proocol is evaluaed hrough boh simulaions and experimens. Boh roueflapping and conrol packe collisions are described, and soluions for hese problems are proposed. The uning of rouing proocol parameers in order o improve he end-o-end conneciviy is sudied in [15]. A performance meric called Rouing Change Laency (RCL) is defined and analyzed. This meric is defined as he ime needed o deermine a new roue afer a link failure, bu i also comprises a ime lapse afer he new roue is discovered, unil i is acually used. This ime lapse is denoed as T new_roue. I is no explained, bu observed o vary beween 4.62 s and 8.86 s

10 7. Conclusions and furher work The rerouing ime is an imporan performance measure in MANETs where node mobiliy is usually high, and conneciviy beween nodes may be disruped frequenly. For ongoing daa raffic ha suffers from link failures, i is highly desirable o reesablish conneciviy hrough alernaive pahs as fas as possible. In his paper we have looked closer on a simple scenario where we have idenified ha queueing is among he main facors having considerable impac on he rerouing ime. The laency relaed o queueing is mainly affeced by wo parameers, namely he ransmi queue size and he rery limi. A large ransmi queue size may resul in a oo high amoun of garbage packes wih sale rouing informaion being insered ino i. In addiion, a high rery value may resul in oo many wased reransmission aemps for hese garbage packes. The combinaion of hese facors migh exend he rerouing ime considerably. We have derived a simple model ha can be used o esimae he rerouing ime. Comparisons of he esimaed and simulaed rerouing imes have shown ha he model is a good approximaion. The analysis is used o explain how queueing migh increase he rerouing ime. In order o solve his problem, we have proposed a simple bu very effecive soluion based on adapive rery limi in he DCF MAC. The queueing problem is resolved by decremening he maximum rery value when successive packes for he same MAC desinaion are discarded due o expiraion of he rery limi. The proposed soluion was implemened and esed in simulaions, and he resuls have shown how effecive i can be. In fac, as long as he daa rae ino he queue is safely below he capaciy of he MAC, he soluion eliminaes he queueing problem associaed wih he rerouing ime. Alhough he proposed soluion seems o be very effecive, here migh be some problems associaed wih i. For example, he soluion migh lead o an unfair resource disribuion in erms of collision avoidance. This needs o be explored in deail, and will be addressed by a follow-on publicaion. I migh also be possible o implemen more complex soluions where he MAC layer discards packes wihou aemping o ransmi hem. Various variaions of our soluion will also be sudied. The proposed soluion is a simple way o resolve queueing relaed delays. We believe here are oher possibiliies in solving he problem or improving he exising soluion. A soluion based on cross-layering, where L2 can send a noificaion up o L3, helping he rouing proocol o deec link breaks much earlier is an exciing area. All his is also lef o fuure works. 8. Acknowledgmen This work was suppored by he ITEA Easy Wireless and CELTIC DeHiGae projecs. References [1] IETF working group Mobile Ad-hoc Neworks, hp:// [2] T. Clausen, P. Jacque, Opimized Link Sae Rouing Proocol (OLSR), RFC 3626, Ocober [3] hp:// [4] IETF MANET OSPF design eam reposiory, hp://hipserver.mc.phanomworks.org/ief/ospf/ [5] ANSI/IEEE Sd , 1999 Ediion (R2003). [6] F. Tobagi and L. Kleinrock, Packe Swiching in Radio Channels: Par 2-The Hidden Node Problem in Carrier Sense Muliple Access Modes and he Busy Tone Soluion, IEEE Trans. Comm., vol. 23, no. 12, pp , [7] G. Chuanxiong, Z. Shaoren, Analysis and evaluaion of he TCP/IP proocol sack of LINUX, Inernaional Conference on Communicaion Technology Proceedings, WCC - ICCT 2000, Vol. 1 (2000), pp vol.1. [8] Nework Simulaor ns-2, hp:// [9] ANSI/IEEE Sd b, 1999 Ediion (R2003). [10] MANET Simulaion and Implemenaion a he Universiy of Murcia (MASIMUM), hp://masimum.dif.um.es/ [11] Engelsad, P.E., Øserbø, O.N., "Analysis of he Toal Delay of IEEE e EDCA and DCF", Proceedings of IEEE Inernaional Conference on Communicaion (ICC'2006), Isanbul, June 11-15, (See also: hp://folk.uio.no/paalee) [12] R. Braden (Edior), Requiremens for Inerne Hoss Communicaion Layers, RFC 1122, Ocober [13] M. Voorhaen, C. Blondia, Analyzing he Impac of Neighbor Sensing on he Performance of he OLSR proocol, Proceedings of he 4h Inernaional Symposium on Modeling and Opimizaion in Mobile, Ad Hoc and Wireless Neworks, April [14] T. Clausen, G. Hansen, L. Chrisensen, G. Behrmann, The Opimized Link Sae Rouing Proocol Evaluaion hrough Experimens and Simulaion, IEEE Symposium on Wireless Personal Mobile Communicaions, Sepember [15] C. Gomez, D. Garcia, J. Paradells, Improving Performance of a Real Ad-hoc Nework by Tuning OLSR Parameers, Proceedings of he 10h IEEE Symposium on Compuers and Communicaions (ISCC), 2005.