h r

arc

ordi GIBBT,

Available online 18 April 2008

lity o

networks where wireless nodes are able to dynamically switch among several link rates.This allows nodes to select the highest possible transmission rate for exchanging data,independently for each neighbor.

2008 Elsevier B.V. All rights reserved.

sioning a difcult problem.Due to the shared media and multihop characteristics of

AWNs, it is known that its capacity can be surprisingly low[1]. Consequently, congestion may easily occur, provokinglosses and high end-to-end delays. In order to avoid con-

completely on the network layer, so that no modicationson lower layers are required, although some informationabout the network congestion state could also be takeninto account if provided by the MAC (medium access con-trol) layer.

Our mechanism takes into account the multirate capa-bility of wireless networks, i.e., it considers that wirelessnodes are able to choose among several modulation

1570-8705/$ - see front matter 2008 Elsevier B.V. All rights reserved.

* Corresponding author. Tel.: +34 934054097; fax: +34 934017055.E-mail address: rafael.guimaraes@ac.upc.edu (R. Guimares).

Ad Hoc Networks 7 (2009) 388400

Contents lists available at ScienceDirect

Ad Hoc Ne

.e lsdoi:10.1016/j.adhoc.2008.04.002Over the last years, ad hoc wireless networks (AWNs),have captured the attention of the research community.The exibility and cost savings they provide, due to the factthat no infrastructure is needed to deploy a AWN, is one ofthe most attractive possibilities of this technology. How-ever, along with the exibility, lots of problems arise dueto the bad quality of transmission media, the scarcity of re-sources, etc.

Since real-time communications will be common inAWNs, there has been an increasing motivation on theintroduction of quality of service (QoS) in such networks.However, many characteristics of AWNs make QoS provi-

sonable solution. However, most of the QoS approachesfound in literature for AWNs do not use reservations. Onereason for that, is the difculty on determining the avail-able bandwidth at a node. This is needed to decide whetherthere are enough resources to accommodate a newconnection.

In this paper we propose a simple, yet effective methodto compute the available bandwidth at a node in AWNs.We use this method to propose a reservation based QoSmechanism. Our proposal not only guarantees certainQoS levels, but also naturally distributes the trafc moreevenly among network nodes (i.e. load balancing). It worksKeywords:Ad hoc wireless networksQuality of serviceRoutingMultirate

1. Introduction gestion, a reservation mechanism that works together witha connection admission control (CAC) seems to be a rea-Received in revised form 19 February 2008Accepted 13 April 2008

topic in the last years. In this paper we describe a QoS reservation mechanism forMultirateAWNs that allows bandwidth allocation on a per ow basis. By multirate we refer to thoseQuality of service through bandwidtwireless networks

Rafael Guimares a,*, Llorenc Cerd a, Jos M. BChris Blondia b

aDepartment of Computer Architecture, Polytechnic University of Catalonia, JbDepartment of Mathematics and Computer Science, University of Antwerp

a r t i c l e i n f o

Article history:Received 22 October 2007

a b s t r a c t

Achieving QoS (qua

journal homepage: wwweservation on multirate ad hoc

el a, Jorge Garca a, Michael Voorhaen b,

irona 1-3, E-08034 Barcelona, SpainMiddelheimlaan 1, B-2020 Antwerpen, Belgium

f service) in ad hoc wireless networks (AWNs) has been a research

tworks

evier .com/locate /adhoc

ter TDMA (Time Division Multiple Access) [12], try to pro-

R. Guimares et al. / Ad Hoc Networks 7 (2009) 388400 389schemes, providing different transmission rates, in order toaccommodate to different channel conditions. We providea set of QoS constraints that must be satised for the ongo-ing QoS ows to consume an overall bandwidth at anynode smaller than or equal to a certain threshold. Alongthe paper we shall refer to this threshold as Q. It may beunderstood as the percentage of time that the channelcan be busy at any given node, because it is transmitting,receiving or listening to trafc that belongs to QoS ows.We propose a set of CAC rules that, upon the assumptionslisted in the following subsection, can satisfy the QoSconstraints.

Finally, we apply our reservation scheme to the opti-mized link state routing protocol (OLSR) [2] although it couldbe applied to other ad hoc routing protocols as well (see [3]for a reference on how to apply such a mechanism to the adhoc on-demand distance vector routing protocol AODV[4]). In [5,6], we have presented preliminary studies ofthe protocol. The results show the feasibility of our schemefor guaranteeing the QoS requirements of accepted ows.

1.1. Our proposal

We treat the problem of achieving end-to-end band-width reservation. Our mechanism, which we call BRAWN(bandwidth reservation over ad hoc wireless networks), isbased on the computation of the available bandwidth seenby a given node and the use of this value to verify whethernew ows can still be routed through this node.

Our scheme is based on the following assumptions: (i)QoS-aware applications are able to request the appropriatebandwidth when establishing a connection. (ii) The nodesknow the capacity of the wireless links that are availablefor QoS ows. Besides this, we assume that the MAC usedis able to isolate trafc classes, in such a way that QoS traf-c has priority over non-QoS trafc (we could, for instance,use 802.11e). This allows nodes to x the previously intro-duced Q threshold. (iii) A pure Carrier Sensing Medium Ac-cess (CSMA) protocol is used. Thus, whenever a node istransmitting, all its neighbors will remain silent. Throughthe paper we shall refer as neighbors each pair of nodesthat are in the receiving range of each other. Note thatwe are not considering a MAC using RTS/CTS, although itcould be easily supported, as we proposed in [3]. (iv) Nodesare able to reach all their neighbors through broadcastpackets.

Of course, the previously described assumptions are notexact in real wireless networks. For instance, the availablecapacity for QoS trafc may be inuenced by non-QoS traf-c and other network conditions. To cope with that, a con-servative value shall be used for Q, or it may be madeadaptive, as we proposed in [7]. Furthermore, changes inthe network conditions, which can be very frequent inAWNs, make the information used by nodes to computethe available bandwidth to be uncertain. Therefore, aftera ow is accepted, its QoS parameters (end-to-end delay,packet loss, etc.) should be constantly monitored in orderto react to congestion. This could be done by re-routingor even dropping some of the involved ows. We will notdeal with these issues, in order to keep the paper focusedon the reservation mechanism.vide a more deterministic solution through the allocationof slots of time for the transmission of nodes. In clusterTDMA, the network is divided into clusters that are coordi-nated by a node elected to be the so called cluster-head.This node is responsible for controlling the intra-clustercommunication by allocating slots of time among thenodes of the cluster.

Another different approach is to provide a solutionpurely on the network layer. These proposals are usuallyindependent of the MAC layer actually used and, thus,can be combined with different layer two mechanisms.In-band signaling support for QoS in mobile ad hoc networks(INSIGNIA) [13], for example, is an in-band signaling proto-col designed explicitly for AWNs. The signaling informa-tion related to the QoS mechanism is encapsulated indata packets, making this approach easy and lightweight.It supports ow reservation, restoration and adaptationalgorithms through the use of control signals carried onan IP option in every data packet.

Other proposals, like the exible quality of service mod-el for MANETs (FQMM) [14], combine a reservation mech-anism for high-priority trafc and service differentiationfor low-priority data. However, this hybrid provisioningscheme does not take into account the characteristics ofad hoc networks.Note that a reservation mechanism approach is moreappropriate for wireless ad hoc networks with xed nodes(e.g. wireless mesh networks [8]) or where mobility is notvery high (e.g. pedestrian networks). If nodes constantlymove with high speeds (vehicular networks, for instance),changes on the topology are very frequent, thus, the re-served path should be constantly updated. For this reasonwe use the term AWN (ad hoc wireless networks) and notMANET (mobile ad hoc networks). MANET is commonlyused in literature to remark the mobility characteristic ofthe AWN under consideration.

The paper is organized as follows. In the next sectionwe present an overview of previous works that also dealwith QoS in AWNs. In Section 3, we rst make an analysison the computation of the amount of bandwidth availablefor QoS reservations. The obtained results are then used topresent the basis of our mechanism in Section 4. In Sec-tion 5 we present an scenario that exemplies howBRAWN actually works. In Section 6 we comment themost relevant aspects of the integration of the BRAWNmechanism into the OLSR protocol. This implementationis then used in some simulations, whose results are shownin Section 7. Finally, we present some conclusions inSection 8.

2. Related work

Many QoS architectures have been proposed for ad hocwireless networks (see [9] and references herein). Some ofthem focus only on the MAC layer, like 802.11e [10] whichis an extension of the original 802.11 standard [11] thatallows the classication of data into several trafc classes,giving more transmission opportunity for data that belongsto higher priority trafc classes. Other proposals, like clus-

sidered to be cross-layer, since they deal with both the

390 R. Guimares et al. / Ad Hoc Networks 7 (2009) 388400network and the link layers at the same time. Some ofthese proposals only gather information from the link layerto be used on routing while others present a more coupledview of both layers. One of the most well known mecha-nisms of this type is SWAN [16]. It is based on the DiffServidea, i.e., on the classication of trafc into classes that areserved with different priorities by the network. By measur-ing MAC delays, it automatically congures a rate controlmechanism and, by measuring the rate of real-time owsthat passes through the neighbors, it evaluates the amountof bandwidth that is still available for new real-timeconnections.

The courtesy piggybacking [17] also proposes a servicedifferentiation solution. It focuses on trying to avoid thebandwidth starvation suffered by low priority trafc whenhigh priority trafc is intense. The idea of this proposal isto piggyback low priority trafc into the high priority pack-ets whenever there is a free space, i.e, whenever a MACframe is not completely lled by the high priority data. In-tense coordination between the MAC and network layer isneeded, since the MAC layer may request low priority datato the network layer in order to ll its frames.

Other solutions, such as the proactive real-time MAC(RTMAC) [18] and the QoS routing protocol proposed by[19], also propose a cross-layer approach. However, theyare both focused on very particular MAC protocols.

None of these proposals, however, deals with reserva-tion of resources in an ad hoc network. Most of them arebased on the idea of providing preferential access to higherpriority trafc classes (or ows). To the best of our knowl-edge, resource reservation for MANETs was rst studied byCansever et al. in [20]. In this paper the authors discuss thedifculty on determining the available bandwidth at anode in a wireless media. The authors give a solution tothis problem for a simple scenario where all nodes usethe same transmission rate, and the topology of the nodesis subject to some restrictions.

The ad hoc QoS on-demand routing (AQOR) proposal[21] introduces a resource reservation-based routing andsignaling algorithm that tries to provide end-to-end qual-ity of service support, in terms of bandwidth and end-to-end delay. AQOR is the proposal we have found that mostresembles our solution. Nevertheless, AQOR does not takeinto account the multirate capability of current networks.A further comparison of our mechanism (BRAWN) andAQOR is presented in Section 5.

3. How much bandwidth is available for reservations?

The BRAWN mechanism is based on the computationof the available bandwidth (AB) by each node in thenetwork in a distributed way. By knowing its availablebandwidth, a node is able to accept or reject a newAnother example of a network layer solution is the core-extraction distributed ad hoc routing algorithm (CEDAR) [15],which is a protocol proposed to reduce the control over-head by dening a backbone that is responsible for allthe route computation on the network.

Finally, some proposals provide solutions that are con-based (CSMA-like) protocol for the analysis that we willpresent throughout the paper.

In order to know howmuch bandwidth is available for anode to use, we must take into account all transmissionsthat directly affect its opportunities to transmit. In the caseof a CSMA-based wireless MAC protocol, the bandwidth ofa node is consumed whenever:

Case 1. It transmits data to a neighbor.Case 2. One of its neighbors is transmitting data (if the

node senses that the medium is being used, itremains in silence).

Representing this in an analytical way, we may statethat the load impact of all transmissions on a node i is gi-ven by

li xi|{z}Case 1

[[

j2Nixj

|{z}Case 2

[j2N

i

xj

; 1

where

li is the load impact of all transmissions (in bps) on nodei.

xi is the total trafc (in bps) that node i wants to trans-mit (either if node i is the source of the trafc or if it isjust forwarding).

Ni is the set of neighbors of node i.

Ni is the set of neighbors of node i and node i itself.

The union operator [ represents a time-based union,i.e., intersections represent parts of the transmissionsthat take place simultaneously. See Fig. 1 for an exampleof the appliance of this operator over two transmissionsthat overlap in time.

The formula derived before can be generalized forwireless multirate networks, i.e., networks where nodescan communicate with each other at different transmissionrates, depending on the wireless medium conditions. To doso, all these values that were represented in bps abovereservation. So, the rst step we should take in orderto dene our mechanism is to compute the AB of eachnode.

If we want to compute the amount of bandwidth that isavailable for a given node to use for new reservations, weshould rst investigate the amount of bandwidth that is al-ready being consumed by active ows. By knowing this va-lue, we may just subtract it from the total bandwidthdedicated to QoS trafc in order to obtain the currentlyavailable bandwidth.

The rst issue that we should notice is that a transmis-sion between two nodes does not consume bandwidthonly from these nodes but also from the whole neighbor-hood, since no other neighbor is able to transmit at thesame time (at least using the same channel) in order toavoid collisions. In fact, the exact knowledge of whichnodes suffer from the interference of a given transmissiondepends directly on the MAC protocol that is being used.For this reason, we assume the use of a carrier sensing

must be normalized, dividing them by the transmission

to implement, some simplications must be done. The rst

Transmission 1

t1

t2t3t4

Transmission 2t4 t1 t2= U

t3 t1 t2=Node A

Node B

Fig. 1. Example of time-based union and intersection operators.

R. Guimares et al. / Ad Hoc Networks 7 (2009) 388400 391of the ad hoc network and after dening the amount ofnormalized bandwidth dedicated to QoS trafc as Q, weare able to state the following QoS constraint that shouldbe respected in order to provide QoS guarantees for real-time ows:

Li 6 Q ; 8i 2S 0 6 Q 6 1; 3rate used:

Li [

j2Ni;8kxjk=vjk

; 2

where

Li is the normalized load impact on node i. From now on,we will consider in this paper only the multirate case(since the single-rate can be seen as a particular caseof a multirate network, where all transmission ratesare the same). We shall use capital letters for referringto normalized values.

xjk is the total trafc (in bps) that node j wants to trans-mit to node k.

vjk is the transmission rate used between nodes j and k.

Since the equation is normalized, if the node is not over-loaded, Li should be a value between 0 and 1.

The use of the union operator states that some trans-missions in the neighborhood may overlap in time. Thiscan happen in CSMA-based networks whenever thesetransmissions do not interfere with each other, as shownby Fig. 2. In this example, transmissions a and b can over-lap in time.

Once we have computed the load impact on each nodei j khg

a b

Fig. 2. Simultaneous transmissions in the neighborhood of node i.of them is related to the computation of the load impact oneach node of the ad hoc network. The use of the unionoperator, as shown by Eq. (2), is not possible, since a nodehas no idea of the degree of simultaneity of the transmis-sions on the neighborhood. For this reason, we simplify theequation by using a simple sum instead, since it is alwaysmore restrictive than using the union (Fig. 1), what stillguarantees the QoS requirements. Thus, the load on a nodei will be computed as

Li [

j2Ni

Xj

Xj2N

i

Xj; 4

where

Xj X8k

xjkvjk

5

is the normalized amount of QoS trafc that node j wantsto transmit (either if node j is the source of the trafc orif it is just forwarding). In BRAWN each node would reservebandwidth for this trafc, thus, Xj can also be interpretedas the total reserved bandwidth at node j. Using a sum inwhereS is the set of nodes that are transmitting or receiv-ing QoS trafc, i.e. the nodes having at least one QoS reser-vation. In the rest of the paper we shall refer S as the QoSset. By guaranteeing condition (3), we can guarantee thatthe channel occupancy due to the QoS trafc observed byany node of the QoS set is never greater than Q. This con-dition should guarantee that there is enough capacity toaccommodate all QoS ows.

Note that Q can be understood as the percentage of timethat the channel can be busy at any node, because either itis transmitting or receiving trafc that belongs to the QoSows. We shall assume that the MAC is able to restrictnon-QoS trafc, such that the normalized capacity Q willbe always available for QoS trafc. This could be achievede.g. using 802.11e, or 802.11 with some additional mecha-nism, e.g. SWAN [16], that regulates non-QoS trafc. Ofcourse, due to collisions, impact of non-QoS trafc andother reasons, the amount of normalized capacity Q avail-able for QoS trafc may vary. To cope with that, a conser-vative value shall be used for Q, or it may be madeadaptive, as we proposed in [7].

4. The basis of BRAWN

As previously mentioned, BRAWN is based on the com-putation of the available bandwidth (AB) in each node ofthe network. The goal of our bandwidth reservation mecha-nism is to provide rate allocation (e.g. peak or sustainablerate) and, at the same time, remain as simple as possible.The solution should provide QoS and yet introduce as littleoverhead as possible in the network. In order to do that, itshould only make use of the information about its 1-hopneighborhood. Since most of the available ad hoc routingprotocols already provide 1-hop signaling, e.g. HELLO mes-sages, any additional information that may be necessarycan be piggybacked on these signaling messages.

In order to provide a simple mechanism that is feasible

Eq. (4) to represent a union may be pessimistic in somescenarios. In [22] the approximation given by Eq. (6) hasbeen proposed

[j2N

i;8kXj

Xj2N

i

Xj X

j;k2Ni

jXj \ Xkj: 6

However, in 802.11-like networks two transmissions can-not overlap in timewhenever either the sender or the recei-ver of one transmission is a neighbor of either the sender orthe receiver of the other one. Therefore, in order to accu-rately compute which intersections from Eq. (6) are notnull, a node would need individual information about everyow in the neighborhood, so that it would be able to iden-tify those that may take place simultaneously. Exchanging

392 R. Guimares et al. / Ad Hoc Networks 7 (2009) 388400this information would introduce too much overhead inthe protocol. Consequently, we have considered Eq. (4) asa convenient approximation for the load demand.

4.1. The available bandwidth in each node

Once each node is able to compute the load demand onitself, this value can be used to establish which part of thetotal bandwidth dedicated to QoS connections is still avail-able for reservations. By using just the information locallyknown by a node (the pre-established Q value and thecomputed load impact), we dene a new value that repre-sents this availability for new ows to be established,which we call the maximum available bandwidth (MAB)

MABi Q Li: 7This value is simply the amount of bandwidth available forQoS ows minus the amount of bandwidth already con-sumed under the point of view of this node, i.e., its load im-pact. By looking at Eq. (3) it is quite simple to notice thatwe may re-write the QoS constraint using this new value

MABi P 0 8i 2S: 8However, knowing the local MAB of a node is not enoughfor the node to decide if new ows can be accepted. Thisis because the available bandwidth of a given node i is alsoaffected by transmissions of its two-hop nodes that haveone of the neighbors of i as a receiver. In Fig. 3, for example,the transmission from g to h only causes an impact on thecomputation of MABg and MABh, although when it takesplace, node i is not allowed to transmit (notice, however,

MABg=0 MABh=0 MABi=1

r=1.0

MAB j=1g h i j

Fig. 3. The maximum available bandwidth and restrictions imposed byneighbors.of the node about the impact of new transmissions on theneighborhood. It is, in fact, the amount of bandwidth avail-able for new transmissions over a given node.

Now, the QoS constraint given by Eq. (8) can be rewrit-ten in terms of the available bandwidth as we state in thefollowing theorem:

Theorem A. Guaranteeing that the AB given by Eq. (9) ofevery node that takes part in a reserved path is non-negative,is equivalent to guaranteeing that the MAB of every node ofthe QoS set is non-negative.

See the proof of this theorem in Appendix A. In otherwords, theQoSconstraintgivenbyEq. (8) canbe rewrittenas

ABi P 0 8i 2 reserved paths: 10Summing up, BRAWN requires that the nodes know thenormalized amount of trafc Xj and the maximum avail-able bandwidth MABj of their neighbors belonging to theQoS set S. These values should be periodically ex-changed among neighbors belonging to S. Nodes that donot belong to S would compute the MABj, which couldbe needed in the CAC of future QoS reservations, but theywould not send it. Each node i uses Xj to compute the loadLi using Eq. (4), and MABi using Eq. (7). Finally, the avail-able bandwidth ABi is computed using Eq. (9).

4.2. Call admission control

After dening the distributed mechanism to computethe available bandwidth at each node of the network, wewill use this value to decide whether a new connectionof r bps ts or not in a given node.

The rst step toward the denition of a call admissioncontrol (CAC) is realizing which transmissions cannot takeplace while a node i is transmitting towards a node j. As wehave discussed before, if we are using a CSMA-like proto-col, none of the is neighbors nor the js neighbors are al-lowed to transmit while i is transmitting to j. Therefore,the following CAC should be checked in every node alonga candidate path:

ABi P[

y2Ni[N

j\path

rvy

; 11

where

i represents the current node in the path.

j represents the next node in the path (to which i willtransmit).

r represents the bandwidth required by the newconnection.

vy is the transmission rate from node y toward its nexthop in the path.that MABi 1). That means that a node also needs to takeinto account its neighbors restrictions.

We, thus, propose to estimate what we call the avail-able bandwidth of a node i (ABi) as the minimum valueof the MABs in its QoS set neighborhood:

ABi minfMABjg; j 2Ni \S: 9This value can also be understood as a more complete view

In this case, just like in the load demand computationEq. (4), we use a simple sum approximation for the unionoperator

ABi PX

y2Ni[N

j\path

rvy

: 12

See the proof that this CAC condition guarantees the QoSconstraint presented by Eq. (10) in Appendix B.

Notice that in the case that nodes move, topologychanges in the network may cause connections that werepreviously accepted by the CAC not to have their QoSrequirements guaranteed after a while. Moreover, even ifQoS can still be guaranteed over a given path, topologychanges may cause more efcient paths to show up, and

before that a new ow rCD 2 Mbps could be accepted at

Table 1Parameters computed by nodes using BRAWN. With ow rAF (a), and withows rAF and rCD (b)

MNA MNB MNC MND MNE MNF

(a) Xi 0.2 0.2 0.0 0.0 0.2 0.0MABi 0.6 0.4 0.6 1.0 0.6 0.8ABi 0.4 0.4 0.4 1.0 0.4 0.6

(b) Xi 0.2 0.2 0.4 0.0 0.2 0.0MABi 0.6 0.0 0.2 0.6 0.2 0.8ABi 0.0 0.0 0.0 0.2 0.0 0.2

Table 2Parameters computed by nodes using AQOR. With ow rAF (a), and withows rAF and rCD (b)

MNA MNB MNC MND MNE MNF

(a) Bself (Mbps) 1 2 0 0 2 1Bavailable (Mbps) 3 2 1 5 2 3

(b) Bself (Mbps) 1 2 2 2 2 1Bavailable (Mbps) 3 0 1 3 0 3

R. Guimares et al. / Ad Hoc Networks 7 (2009) 388400 393being able to use them may optimize the use of networkresources. Therefore, in the presence of movement, theQoS mechanism should be adaptive. This could be achievede.g. by periodically refreshing reservations, so that the net-work is constantly re-validating the admission control andsearching for better routes for previously establishedconnections.

5. Exemplifying BRAWNs behavior

In order to better understand the behavior of theBRAWN mechanism, we will take a step-by-step look atthe ad hoc network example depicted by Fig. 4. In this sim-ple example all links between mobile nodes are 5 Mbps.Assume that in this network there is an established reser-vation for a QoS ow of 1 Mbps following the pathMNA !MNB !MNE !MNF . For simplifying the example,we shall also assume that the reserved capacity for QoStrafc is Q 1.

The row Xi in Table 1 (a) shows the normalized amountof trafc that would be advertised by the nodes. Uponreceiving theses values, each node would compute theMABi shown in the corresponding row of the table. For in-stance, MNB would receive XA 0:2;XC 0:0 and XE 0:2.Since XB 0:2, it would compute MABB 0:4. Finally, uponreceiving the MAB from their neighbors, nodes would com-pute the ABi given in the table. Note that nodes MNC andMND would not advertise their MAB, because they do notbelong to the QoS set.

Assume that after this, node MNC wishes to establish anew QoS ow of rCD 2 Mbps with node MND. The follow-

MNA MNB MNC MND

MNE

MNF

vAB = 5 Mbps vBC= 5 Mbps vCD= 5 Mbps

vEF = 5 MbpsrAF = 1 Mbps

vBE = 5 Mbps vCE= 5 Mbps

Fig. 4. Network topology.ing CAC conditions would be checked: ABC P 0:4 andABD P 0:4 (2 Mbps/5 Mbps = 0.4). Thus, the ow wouldbe accepted, and the values of Xi, MABi and ABi wouldchange as shown in Table 1(b).

We may intuitively check that, after accepting the owrCD, the available bandwidth computed by the nodes is cor-rect: Whenever one of the nodes MNA;MNB;MNC and MNEsends a packet, all the others in this set must remain silent.Since altogether send 5 Mbps, which is the link capacity,their available bandwidth is 0. Node MNE must be silentwhenever MNB;MNC or MNF transmit. Since nodes MNE,MNB and MNC transmit altogether 4 Mbps, the availablebandwidth at node MNF is 1 4/5 = 0.2. Similarly, we canderive that the available bandwidth at node MND is also0.2.

5.1. Comparing BRAWN to AQOR

Among the previously proposed protocols, AQOR is thesolution that most resembles that of BRAWN. In this sec-tion we use the previous example to compare BRAWNand AQOR in terms of the calculation of the availablebandwidth.

In AQOR the authors dene BselfI as the total trafctransmitted or received at a node I. Bself is periodically ex-changed between neighbors. Then, the available band-width (Bavailable) is computed as

BavailableI BXJ2NI

BselfJ; 13

where B is the maximum transmission bandwidth (5 Mbpsin the above example), and NI is the neighborhood ofnode I. Table 2 shows the Bself and Bavailable values thatwould be computed by the nodes using AQOR in Fig. 4(we did not use normalized values, as it was done inBRAWN, since AQOR was not dened for multirate net-works). Note that AQOR would estimate an available band-width of only 1 Mbps at node MNC , while we have seen

node MNC . Nevertheless, if the ow rCD were accepted, thenodes would update Bself and Bavailable as shown in Table 2(b). Note that the value Bself would become negative atnode MNC , meaning that there has been an over-reserva-tion of resources. In fact, the authors of AQOR have re-ported that the more trafc is sent in the neighborhood,the more conservative is the estimation of the availablebandwidth. Therefore, we conclude that BRAWN is ableto estimate the available bandwidth much moreaccurately.

6. Implementation issues

BRAWN can be integrated into many routing protocolsproposed for AWNs. In this section we explain how wehave integrated it in one protocol that uses a proactive ap-proach: OLSR [2].

OLSR provides a neighbor discovery mechanism basedon the periodic broadcast of HELLO messages. These mes-sages are broadcasted to the one-hop neighborhood and,by receiving them, a node is able to be aware of its neigh-bors. BRAWN makes use of these messages by piggybac-king on them the information that a node should haveabout its neighbors (Xj and MABj) in order to compute

(iii) OLSR TC messages are modied to also advertise ABiand vij of each of node is MPR selectors. By doingthat, each node has knowledge of the network topol-ogy and the bandwidth available in the network.

(iv) In order to nd a route that meets the QoS require-ments, we modied the OLSR route selection algo-rithm to nd a shortest hop path that has enoughbandwidth (the shortest-widest path) to meet theserequirements. Since TC messages also advertise ABiand vij, the originating node has sufcient informa-tion to decide if enough resources are available(see Eq. (12)).

(v) Bandwidth reservation at intermediate nodes isdone through the exchange of Reservation Request/Reservation Reply messages previously to sendingdata packets.

6.1. The CAC in OLSR

The OLSR routing protocol uses an optimized version ofthe Dijkstra algorithm to compute shortest path routes toall the other nodes in the network. However the QoS rout-ing in BRAWN needs to nd a route towards a specic des-

the CA

394 R. Guimares et al. / Ad Hoc Networks 7 (2009) 388400Fig. 5. Integration ofthe load impact Li and the available bandwidth ABi, asseen in Eqs. (4) and (9).

The detailed changes required to integrate BRAWN intoOLSR are the following:

(i) OLSR HELLO messages are modied such that eachnode i with QoS reservations advertises Xi andMABi to their neighbors. This is all the informationthat node is neighbors need to compute Eq. (4).

(ii) Each node i collects these QoS HELLO messages fromtheir neighbors to compute ABi according to Eq. (9).tination that meets the bandwidth requirements of thenew ow and at the same time avoids a QoS violationshould the ow be allowed. Both goals are accomplishedby applying the CAC algorithm as shown in Section 4.2.

We modied the default route selection algorithm fromOLSR so that it is able to compute a route for QoS ows thatmeets bandwidth requirements and delivers a shortestwidest path. The CAC is performed on each new link thatis added to the nodes topology map so that links thatwould result in the QoS constraint being broken some-where along the route will no longer be taken into accountby the routing algorithm. As previously explained, each

C algorithm in OLSR.

node gathers from HELLO and TC packets the necessaryinformation to perform the CAC.

Fig. 5 shows a pseudo code for the route computationalgorithm. It is based on the route computation algorithmdened by OLSR, to which we added especially importantparts for the QoS routing decisions (in uppercase). Whilethis algorithm can compute shortest-widest paths to allother nodes in the network, we are only interested inone to the requested destination. After this route is found,the reservation signaling reserves resources along thepath.

6.2. Reservation signaling interaction

Reservation of requested bandwidth is done by sendinga reservation request (ResvReq) from the source towards

7. Simulation results

We have added our reservation scheme in an OLSRimplementation and then simulated its behavior usingns-2 [23] version 2.29. The code that we have added tons-2 and used in this paper is available at http://www.pat-s.ua.ac.be/software/brawn, together with example scripts,so that simulations may be easily reproduced.

Simulations were run using the scenario described asfollows:

MAC: 802.11 with multirate.

Multirate parameters: for a distance less than 50 m therate is 11 Mbps; for a distance between 50 and 90 mthe rate is 2 Mbps, according to the ORiNOCO 802.11bPC card specication for a semi-open environment [24].

vReq

vRep

rvatio

R. Guimares et al. / Ad Hoc Networks 7 (2009) 388400 395h i j

ResvReq ResvReq Res

...

ResvRep ResvRep Res

ActivateFlow

ActivateFlow

ActivateFlow

Fig. 6. Resethe destination (Fig. 6). On receiving such a request anintermediate node determines the next hop for the QoSow, which also involves checking that the QoS constraintis not being violated.

If the destination is reached, it sends back a positive re-ply towards the source using the reverse path. On seeingthe reply coming through, the intermediate nodes installsthe QoS route into their active routing table.

Should it occur that the QoS constraint would be vio-lated in one of the intermediate nodes, then a negative re-ply is returned, containing new information on theavailability of bandwidth at this intermediate node. Usingthis information the source can try to reserve resourcesalong another path, if it exists.

By performing the reservation this way, we still stayclose to the OLSR philosophy, where the link state routingalgorithm is only used to nd a suitable next hop towardsthe destination. On the one hand the source can use theCAC to locally decide if it is necessary to block the ow,on the other hand all nodes work together in a distributedmanner to decide the best path towards the destination.The idea behind this is that, although all nodes in thenetwork obtain information about the topology andthe available bandwidth, this information might not beup to date since the topology signaling is onlyperformed periodically and also happens at a lower ratethan the HELLO signaling related to the localneighborhood.

CAC CAC CACk l m ...

ResvRep ResvRep

ActivateFlow

ActivateFlow

ActivateFlow

n signaling.Using the following parameters for the OLSR protocol:

HELLO_INTERVAL: 1 s.

TC_INTERVAL: 2 s.

NEIGHB_HOLD_TIME: 5 HELLO_INTERVAL TOP_HOLD_TIME: 3 TC_INTERVAL DUP_HOLD_TIME: 25 s.

The rst evaluation we should do is the election of thevalue for the QoS parameter Q, since it should not be cho-sen arbitrarily. Using a simulation setup with 40 nodes andvarying Q from 10% to 22.5% (0.10.225) in steps of 1.25%,we have investigated the effect of this parameter on theend-to-end delay and packet loss measurements. Note that

ResvReq ResvReq

CAC CAC CAC Carrier sensing range: 200 m (around 2.2 2 Mbpstransmission range).

CBR connections sending 500 bytes packets with a32 kbps rate.

2070 nodes randomly placed over a square of300 300 m.

20 ows are initiated (one each 15 s) between randompairs of nodes.

The simulation time is 400 s, including a 40 s startupperiod that gives OLSR the time to exchange routinginformation before applications start.

Nodes do not move.

we have run 10 simulations with different node positionsfor each value of Q, so that the results shown below repre-sent the average value with a condence interval.

Results are summarized in Figs. 7 and 8: the former pre-sents the average 99.9 end-to-end delay percentile of theows for different values of Q while the latter presentsthe percentage of packets that were lost.

Here Q should express the maximum utilization of thewireless medium by QoS trafc for which end-to-end de-lays are still bounded (preferred to be lower than 150 msfor interactive multimedia applications [25]). Note thatthis utilization (12.5%) lies well beneath the theoreticalmaximum of 1/3 [26,27], which is to be expected in ran-dom scenarios. However, the performance of the link layercan still be affected by many aspects (e.g. carrier sensingrange) and as such the value of Q for which we still achievethese desirable end-to-end delays can also differ. Q thus of-fers a trade off between accuracy and the complexity of themodel. For the following simulations, we have chosen Q tobe equal to 12.5%. As mentioned earlier, it is also possibleto make an adaptive version of the protocol in which Q isdynamically matched to the network potential.

Having chosen a value for the Q parameter, we havecompared the BRAWN mechanism to a scenario where no

QoS is supported (using the standard OLSR protocol). Byvarying the network density, we were able to observe thenumber of ows that were accepted by the CAC of theBRAWN mechanism and the number of ows that couldbe accepted by OLSR before the network nodes reachedan occupancy of 15% (about the same one allowed byBRAWN for Q 12:5%, when considering the overheadgenerated by IP and ethernet headers, by OLSR and bythe reservation signaling). See Fig. 9 for these results.

As we have previously discussed, BRAWN naturally dis-tributes the trafc more evenly through the network, sincewhenever a node is about to reach congestion, it is no long-er used in new paths. That makes new ows to be routedthrough (possibly longer) paths that avoid potential con-gested areas. Due to this behavior, BRAWN is able to acceptmore ows than standard OLSR when using this occu-pancy cut-off.

This cut-off, however, was articially introduced intoOLSR in order to compare the load-balancing that is pro-vided by our mechanism. In its standard behavior, OLSRnever rejects new ows, even when the network is alreadycongested. This behavior can be easily seen in Figs. 10 and11.

200

300

400

500

600

700

800

axim

um o

f 99.

9 pe

rcen

tile

o

ver

all

flow

s. (m

s)

BRAWN

2

4

6

8

10

12

# of

app

licat

ions

star

ted

OLSRBRAWN

396 R. Guimares et al. / Ad Hoc Networks 7 (2009) 3884000

100

0.1 0.125 0.15 0.175 0.2 0.225

M

Q

Fig. 7. 99.9 end-to-end delay percentile for different values of Q.

0

0.1

0.2

0.3

0.4

0.5

0.1 0.125 0.15 0.175 0.2 0.225

% o

f pac

kets

lost

Q

BRAWN

Fig. 8. Packet loss for different values of Q. 0 20 30 40 50 60 70

# of nodes

Fig. 9. Number of accepted ows for different network densities.

0

5

10

15

20

25

30

20 30 40 50 60 70Max

imum

of 9

9.99

per

cent

ile o

ver a

ll flo

ws.

# of nodes

OLSRBRAWN

Fig. 10. 99.9 end-to-end delay percentile (in seconds) for different net-work densities.

We then launched a single scenario where 8 connec-tions were accepted to be analyzed. Fig. 13 shows the evo-lution of the connections that were established when usingeach protocol on this specic scenario. Note that all con-nections are established with OLSR while only eight withBRAWN (the others are blocked).

Following the analysis of the same scenario, we are ableto notice that end-to-end delays and packet loss increasesignicantly around 300 s of simulation. At this momentthe MAC gets congested and requirements can no longerbe guaranteed for the previously established ows whenusing OLSR. Fig. 14 depicts the maximum end-to-end delayof CBR packets, and Fig. 15 depicts the maximum percent-age of packets lost by connections, measured in intervals of1.25 s (the transmission time of 10 packets). These guresshow us that BRAWN is not only successful in avoidingnetwork congestion, but also in avoiding packet lossesand increased delays. Compared to BRAWN, OLSR behavesmuch worse, since it looses up to 50% of the packets atsome instances.

It is also interesting to know how many connections aresuffering from congestion. Figs. 16 and 17 show the trans-mission delay complementary cumulative distribution

0

5

10

15

20

25

30

35

20 30 40 50 60 70

% o

f pac

kets

lost

# of nodes

OLSRBRAWN

Fig. 11. Packet loss probability for different network densities.

0

2

4

6

8

10

12

14

16

18

20

Num

ber o

f Con

nect

ions

OLSRBRAWN

R. Guimares et al. / Ad Hoc Networks 7 (2009) 388400 397The rst gure shows the average value of the 99.9 per-centile end-to-end delay suffered by ows when usingBRAWN or OLSR. With BRAWN, delays are always underan acceptable limit, since the admission of new ows is lim-ited by the mechanism and routes are well spread throughthe network. When using OLSR, however, delays are veryhigh and increase as the network density gets higher.

The second gure shows the packet loss probability. Byusing our mechanism, losses are almost insignicant, whilethe lack of control of standard OLSR leads to high packetloss rates.

In order to provide a better understanding of the behav-ior of the mechanism, below we present some gures thatanalyzes the dynamics of BRAWN on a single run with 40nodes deployed on random positions. Since results on asingle simulation may be inuenced by many random fac-tors (e.g. position of the nodes), we rst repeated the sim-ulation 200 times, each of them using different nodeplacements in order to guarantee that this single simula-tion is a representative scenario, i.e., the number of ac-cepted ows in this single simulation is close to theaverage number of accepted ows in many different simu-lation runs. Fig. 12 shows the results we obtained for these200 repetitions. BRAWN accepts an average of eight con-nections with a standard deviation of one connection.{0}%

{5}%

{10}%

{15}%

{20}%

{25}%

{30}%

3 4 5 6 7 8 9 10 11 12 13

% o

f occ

uren

ce

# accepted connections

Accepted Connections (200 experiments)Gaussian Distribution (u=8, sd=1)

%

Fig. 12. Probability of accepting a given number of connections whenusing BRAWN. 100 200 300 400 500 600

Time (s)Fig. 13. Connection setup.

0.001

0.01

0.1

1

10

100

100 200 300 400 500 600

Del

ay (s

)

Time (s)

OLSR

BRAWN

Fig. 14. Maximum delay.

0.00010

20

40

60

80

100

100 200 300 400 500 600

% o

f pac

kets

lost

in 1

.25

sec

(Tx T

ime o

f 10 p

acke

ts)

Time (s)

OLSRBRAWN

Fig. 15. Maximum loss.

398 R. Guimares et al. / Ad Hoc Networks 7 (2009) 388400function (CCDF), i.e. Probftransmission delay > xg, for allestablished connections with OLSR and BRAWN. Fig. 16shows us that using OLSR, all but two ows have a 10%chance of having at least a 1 s delay. With BRAWN on theother hand the every ow has end-to-end delays smallerthan 100 ms with a probability higher than 99%.

8. Final remarks

In this paper we have described a bandwidth reserva-tion scheme for ad hoc networks that satises the follow-ing QoS constraint: The load demand offered to thewireless media by the QoS trafc observed at any node ina path that is about to be established 6 Q. Parameter Qis dimensioned in a way that delays are acceptable forQoS connections. Our reservation scheme is designed fornetworks where nodes can communicate to neighborsusing different transmission rates depending on channelconditions (multirate ad hoc networks) and only requiresthat nodes know the normalized bandwidth reservationand maximum available bandwidth of their neighbors.These quantities can be easily advertised by means of HEL-LO packets. We also give a CAC rule that nodes should ap-ply to new connections requiring QoS. 0.0001

0.001

0.01

0.1

1

0.001 0.01 0.1 1 10 100

CDPF

Delay (s)

OLSR

Fig. 16. OLSR delay histogram.This work was supported by CAPES Brazil, the Minis-try of Education of Spain under Grant CICYT TSI2007-66869-C02-01, the Fund for Scientic Research Flandersunder Scientic Research Community Broadband commu-nication and multimedia services for mobile users, byDWTC Belgium under Project IAP P5/11 MOTION (MobileWe have described how to integrate our reservationscheme with the OLSR routing protocol and we haveimplemented it using the ns-2 simulator. We have thensimulated OLSR with and without our reservation scheme.The following items summarize our ndings:

Ad hoc networks can easily become congested by QoStrafc (differently from TCP, this kind of trafc typicallydoes not provide congestion control mechanisms).

Congestion can easily extend to most of the networkintroducing high delays and losses, damaging, thus,most of the connections that requires QoS.

Our reservation scheme provides a feasible way to avoidcongestion, guaranteeing, thus, QoS requirements toongoing connections.

Acknowledgments

0.001 0.01 0.1 1 10 100Delay (s)

Fig. 17. BRAWN delay histogram. 0.001 0.01

0.1

1

CDPF

BRAWNMultimedia Communication Systems and Networks), byIWT under Project 020152. End-to-end QoS in IP based Mo-bile Networks, the European NoE EuroFI and the EuropeanProject WIDENS.

Appendix A. Proof of Theorem A

Theorem A. Guaranteeing that the AB given by Eq. (9) ofevery node that takes part in a reserved path is non-negative,is equivalent to guaranteeing that the MAB of every node ofthe QoS set S is non-negative.

Proof. The computation of the MAB of a given node takesinto account only information about transmissions per-formed by the node and its 1-hop neighbors (see Eq. (7)).Consequently, a transmission between two nodes onlyimpacts the MAB of the 1-hop neighborhood of the sender.

X r

R. Guimares et al. / Ad Hoc Networks 7 (2009) 388400 399MABi P 0 8i 2S 17h

Appendix B. Proof that the CAC guarantees the QoSconstraint

The use of the CAC proposed in Eq. (12) guarantees thatthe QoS constraint dened in Eq. (8) is respected in everycondition. In fact, as we demonstrate below, it is guaran-teed in all cases when using a simple sum approximationfor the union operator.

Proof. As we have previously demonstrated, in order toguarantee the QoS constraint presented in Eq. (8), we canlimit ourselves to guaranteeing the condition presented byEq. (10). Thus, we just need to demonstrate that theproposed CAC guarantees that after accepting a new owof r bps, every node in the ow path present a non-negative AB, i.e., all the nodes in the ow path and their 1-hop neighborhood belonging to the QoS set, present a non-negative MAB. Since we are only concerned with the nodesbelonging to the QoS set, in the following we shall refer toonly this set of nodes.

Nodes in the ow path: for a given node i in the path, wewant to guarantee that its MAB is non-negative in themoment t1 just after the acceptance of the new ow (t0represents the moment just before the acceptance).

MABit1P 0() Qi Lit1P 0() Qi Xy2N

i

Xyt1P 0:

Since the only new transmissions from t0 to t1 are the onesdue to the accepted ow, we have:By observing this, we can state that whenever a newreal-time ow is established over the network, only nodesthat takes part in the ow path and their 1-hop neighbor-hood are affected by these new transmissions. All the othernodes throughout the network see no changes on theirMAB. Thus, if they were non-negative before the ow wasestablished, they would remain like that afterwards.

Then:

ABi P 0 8i 2 resvd paths()minj2N

i

fMABjgP 0 8i 2 resvd paths()

MABj P 0 8j 2Ni 8i 2 resvd paths14

So, considering that in the beginning all nodes of the QoSset have non-negative MABs and that nodes that are notin the 1-hop neighborhood of reserved paths do not seeany changes on their MAB, we can conclude that

MABj P 0 8j RNi 8i 2 resvd paths()MABj P 0 8j 2Ni 8i 2 resvd paths

15

By using the results of (14) and (15):

MABj P 0 8j 2 Ni [Ni ; 8i 2 resvd paths()MABi P 0 8i 2S

16So, as we were willing to demonstrate:

ABi P 0 8i 2 resvd paths()ABit0Py2N

i[N

j\path vy

: 19

Since we know that ABit0PMABit0, we may also saythat the CAC guarantees that

MABit0PX

y2Ni[N

j\path

rvy

: 20

And nally, since Eq. (18) is more restrictive than Eq. (20)(note the additional terms in the former, as well as the sumover a more restricted set of nodes), we may say that theCAC satises the desired conditions.

Nodes in the 1-hop neighborhood of the ow path: Thereare basically two different cases that should be taken intoaccount:

1. Node n that is a neighbor of a node i in the path and allits other neighbors in the path are in the neighborhoodof i or j (considering that i transmits to j).

2. Node whose neighbors in the path are more spread.

In the rst case, we have that:

MABnt1P 0()MABnt0PX

y2Nn \path

rvy

: 21

Since we know that

ABit0 6MABnt0Nn \ path Ni [Nj \ pathand since the CAC guarantees Eq. (19), we are also able toguarantee the condition expressed by Eq. (21).

Since transmissions that take place outside the neigh-borhood of the sender and the transmitter may overlap intime, the second case may, in fact, be seen as a combina-tion of several non-correlated occurrences of the rst case.So, if the CAC guarantees the QoS constraints for the rstcase, it will guarantee for the second as well. h

References

[1] P. Gupta, P.R. Kumar, The capacity of wireless networks, IEEETransactions on Information Theory 46 (2) (2000) 388404.

[2] T. Clausen, P. Jacquet, RFC 3626: Optimized Link State RoutingProtocol (OLSR), October 2003.

[3] R. Guimarpes, Ll. Cerda, Bandwidth reservation on wireless networkswith RTS/CTS signalling, in: Proceedings of the 12th EuropeanWireless Conference (EW2006), 2006.

[4] C. Perkins, E. Belding-Royer, S. Daas, RFC 3561: Ad hoc On-DemandDistance Vector Routing, July 2003.

[5] Ll. Cerda, M. Voorhaen, R. Guimarpes, J.M. Barcelo, J. Garcfa, J.Morillo, C. Blondia, A reservation scheme satisfying bandwidth QoSQi X

y2Ni

Xyt0|{z}

Lit0

X

y2Ni\path

rvyP 0

We then have:

MABit0PX

y2Ni\path

rvy

: 18

So, in order to be correct, the CAC must guarantee Eq. (18).Our CAC, in simple sum approximation form, expressed byEq. (12) guarantees the following condition:

400 R. Guimares et al. / Ad Hoc Netconstraints for multirate ad-hoc networks, in: IST Mobile andWireless Communications Summit, 2005.

[6] Ll. Cerda,M. Voorhaen, R. Guimarpes, J.M. Barcelo, J. Garcfa, C. Blondia,A reservation scheme satisfying bandwidthQoS constraints for ad-hocnetworks, LectureNotes in Computer Science (LNCS) 3427 (2005) 176.

[7] R. Guimarpes, Ll. Cerda, Lecture Notes in Computer Science (LNCS),vol. 4396. pp. 102112.

[8] Ian F. Akyildiz, Wang, Xudong, Wang, Weilin, Wireless meshnetworks: a survey, Computer Networks 47 (4) (2005) 445487.

[9] T.B. Reddy, I. Karthigeyan, B.S. Manoj, C.S.R. Murthy, Quality ofservice provisioning in ad hoc wireless networks: a survey of issuesand solutions, Ad Hoc Networks 4 (1) (2006) 83124.

[10] IEEE Std 802.11e/D3.0 (Draft Supplement to IEEE Std 802.11).Medimum Access Control (MAC) Enhancements for Qualiy ofService, May 2002.

[11] ANSI/IEEE Std 802.11, 1999 Edition, Wireless LAN Medium AccessControl (MAC) and Physical Layer (PHY) Specications, 1999.

[12] M. Gerla, J.T.C. Tsai, Multicluster mobile multimedia radio network,Wireless Networks 1 (3) (1995) 255265.

[13] S. Lee, G. Ahn, A.T. Campbell, Improving UDP and TCP performance inmobile ad hoc networks with INSIGNIA, IEEE CommunicationsMagazine 39 (6) (2001) 156165.

[14] H. Xiao, W.G. Seah, A. Lo, K.C. Chua, A exible quality of servicemodel for mobile ad hoc netowrks, in: Proceedings of the IEEEVehicular Technology Conference, 2000.

[15] R. Sivakumar, P. Sinha, V. Bharghavan, CEDAR: a core-extractiondistributed ad hoc routing algorithm, IEEE Journal on Selected Areasin Communications 17 (8) (1999) 14541465.

[16] A. Veres, G. Ahn, A.T. Campbell, L. Sun, SWAN: service differentiationin stateless wireless ad hoc networks, in: Proceedings of theConference on Computer Communications (IEEE Infocom), 2002.

[17] W. Liu, Y. Fang, Courtesy piggybacking: supporting differentiatedservices in multihop mobile ad hoc networks, in: Proceedings ofthe Conference on Computer Communications (IEEE Infocom), 2004.

[18] V. Vivek, T. Sandeep, B.S. Manoj, C. Siva Ram Murthy, A novel out-of-band signaling mechanism for enhanced realtime support in tacticalad hoc wireless networks, in: Proceedings of the IEEE Real-Time andEmbedded Technology and Applications Symposium, 2004.

[19] M. Canales, J.R. Gallego, A. Hernandez, A. Valdovinos, Cross-layerrouting for QoS provision in multiservice mobile ad hoc networks,in: Proceedings of the 17th Annual IEEE International Personal,Indoor and Mobile Radio Communications, 2006.

[20] D.H. Cansever, A.M. Michelson, A.H. Levesque, Quality of servicesupport in mobile ad-hoc IP networks, in: Proceedings of the IEEEMilitary Communications Conference, 1999.

[21] Q. Xue, A. Ganz, Ad hoc QoS on-demand routing (AQOR) in mobile adhoc networks, Journal of Parallel and Distributed Computing (63)(2003) 154165.

[22] Y. Gao, D. Chiu, J.C.S. Lui, Determining the end-to-end throughputcapacity in multi-hop networks: methodology and applications, in:Proceedings of the SIGMetrics/Performance06, 2006.

[23] The Network Simulator ns-2. URL .[24] Datasheet for Orinoco 11b Client PC Card. URL .[25] ITU-T Study Group 12. URL .[26] B.Villela,O.Duarte,Maximumthroughputanalysis inadhocnetworks,

in: Proceedings of the Third IFIP-TC6Networking Conference, 2004.[27] P.C. Ng, S.C. Liew, Throughput Analysis of IEEE 802.11 Multi-hop ad

hoc Networks, to Appear IEEE/ACM Transaction on Networking, June2007.

Rafael Guimares received his ComputerEngeneering degree from the Federal Univer-sity of Esprito Santo (UFES), Vitria, Brazil in2000 and the M.Sc. in Telematics and Tele-communications from the State University ofCampinas (UNICAMP), Campinas, Brazil in2001. He is currently a Ph.D. student inComputer Networking Group of the ComputerArchitecture Department at Polytechnic Uni-versity of Catalonia (UPC), Barcelona, Spain.His research interests include QoS and routingon wireless ad hoc networks.Chris Blondia obtained his Master in Scienceand Ph.D. in Mathematics, oth from the Uni-versity of Ghent (Belgium) in 1977 and 1982respectively. In 1995 he joined the Depart-ment of Mathematics and Computer Scienceof the University of Antwerp (UIA), where heis currently a Professor and head of the Per-formance Analysis of TelecommunicationSystems (PATS) research group. His mainresearch interests are related to mathematicalmodels for performance evaluation of com-puter and communication systems and the

impact of the performance on the architecture of these systems.Michael Voorhaen received the Mastersdegree in Computer Science from the Univer-sity of Antwerp in 2003 with great distinction.In the same year, he joined the PerformanceAnalysis of Telecommunication Systems(PATS) research group at the University ofAntwerp where he is currently working as aresearch assistant. His Ph.D. research includeswork on scalability and QoS in mobile ad hocnetworks. His main research interests are inwireless ad hoc networks, sensor networksand, more generally, in broadband wirelesscommunications.Jorge Garca graduated as TelecomunicationsEngineer in 1988 (UPC), Barcelona, Spain, andholds a Ph.D. in Telecommunications (UPC,1992, Best UPC Telecommunication ThesisAward). During 199293 he was a visitingscientist at University of Arizona (M.F. Neuts).Since 2003 he is full professor at the Com-puter Architecture Department of UPC, wherehe is head of the Computer NetworkingResearch Group. His current research interestsare network mobility, ad hoc and sensor net-works, software implementation of network-

ing protocols and performance evaluation of computer systems.Jos Maria Barcel graduated in Telecommu-nications Engineering from the PolytechnicUniversity of Catalonia (UPC), Barcelona, Spain,in 1991 and got the Ph.D. degree in Telecom-munications Engineering fromUPC in 1998. Hejoined the Computer Architecture Departa-ment in 1993, where he is a full time associateprofessor. His areas of interest are TCP/IP,mobile communications and Internet routing.Llorenc Cerd obtained the engineeringdegree in telecommunications in 1993 fromPolytechnic University of Catalonia (UPC),Barcelona, Spain. He joined the ComputerArchitecture Department of UPC in 1994,where he obtained his PhD degree in 2000.Currently he is Assistant Professor. His areasof interest include TCP, IP micromobility,wireless 802.11 networks, C++ programming,simulation, performance evaluation, queueingnetworks.works 7 (2009) 388400

Quality of service through bandwidth reservation on multirate ad hoc wireless networksIntroductionOur proposal

Related workHow much bandwidth is available for reservations?The basis of BRAWNThe available bandwidth in each nodeCall admission control

Exemplifying BRAWN ' s behaviorComparing BRAWN to AQOR

Implementation issuesThe CAC in OLSRReservation signaling interaction

Simulation resultsFinal remarksAcknowledgmentsProof of Theorem AProof that the CAC guarantees the QoS constraintReferences