Coursehero >> California >> UCLA >> CS 465

Course Hero has millions of student submitted documents similar to the one below including study guides, homework solutions, papers, exam answer keys and textbook solutions.

MANET ProbeCast: Admission Control via Probing Department of Computer Science, University of California, Los Angeles Los Angeles, CA 90095,USA Soon Y. Oh, Gustavo Mar a, and Mario Gerla soonoh@cs.ucla.edu, gmar a@cs.ucla.edu, gerla@cs.ucla.edu ABSTRACT An inelastic ow is a ow with inelastic rate: i.e., the rate is xed, it cannot be dynamically adjusted to tra c and load condition as in elastic ows like TCP. Real time, interactive sessions and video/audio streaming are typical examples of inelastic ows. Reliable support of inelastic ows in wireless ad hoc networks is extremely challenging because ows and routes dynamically change and ows compete for the shared wireless channel. Bandwidth must be reserved for inelastic ows at session set up time. To avoid repeated attempts to set up reservations in a volatile network and prevent serious network capacity degradation due to call set up overhead, a Call Admission Control strategy robust to mobility must be developed. In this paper we propose ProbeCast, a probe based call admission control scheme with QoS guarantees for inelastic ows. ProbCast was designed for multicast streams but can also work, by default, for unicast. In ProbeCast, a path (or a tree) is probed for capacity availability. If an intermediate link along the probed path fails to meet the QoS requirement, the ow is pushed back via backpressure upstream to the source. The backpressure principle is simple; however its implementation requires some care to avoid unfairness and eventually capture by one of the ows sharing a congested bottleneck. We show that proportional fairness among inelastic contenders will prevent capture. To achieve this, we have developed the Neighborhood Proportional Drop (N-PROD) scheme. N-PROD guarantees the same proportional drop rate among all ows competing in the same contention domain. We demonstrate the e cacy and robustness of ProbeCast for unicast as well as multicast scenarios using the Qualnet simulation platform. General Terms Algorithm, Design, Performance Keywords MANET, Call Admission Control, QoS, Inelastic ow 1. INTRODUCTION Categories and Subject Descriptors C.2.1 [Computer-Communication Networks]: Network Architecture and Design Wireless communication ; C.2.2 [Computer-Communication Networks]: Network Protocols Routing Protocols Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for pro t or commercial advantage and that copies bear this notice and the full citation on the rst page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior speci c permission and/or a fee. Q2SWinet 08, October 27 28, 2008, Vancouver, BC, Canada. Copyright 2008 ACM 978-1-60558-237-5/08/10 ...$5.00. In emergency and tactical Mobile Ad Hoc Networks (MANETs), audio and video streaming are essential requirements for group interoperation. Even commercial MANETs (e.g., vehicular networks) will be faced with multimedia streaming because of the popularity of applications such as P2P TV and YouTube. We can expect that future MANETs will be designed to handle inelastic ows, both uni and multicast with QoS (bandwidth and delay) requirements, in addition to elastic ows such as TCP. Most of the previous work on QoS in MANETs is based on resource reservation on a selected path before transmission can commence [3, 4, 5, 11, 13]. Well known limitations of traditional Call Admission Control (CAC) strategies are: the complexity of the available bandwidth estimation in a shared wireless environment (where each allocation impacts several other ows in di erent ways depending on their relative positions and transmission ranges); the volatility of the available bandwidth estimation due to the rapidly changing topology; the need to frequently reallocate resources due to node mobility, and; the overhead introduced by the frequent updates. However, if inelastic calls are accepted without any attempt to allocate resources, the situation is even worse: congestion will set on and no inelastic calls can get through! As a solution to this dilemma, we propose ProbeCast, a probing based CAC scheme for inelastic ows that does resource allocation without PRIOR RESERVATIONS. Namely, a new ow rst probes the availability of resources, e.g., evaluating packet drop rate at intermediate nodes without any notion of resource allocation. If an intermediate link fails to meet the QoS requirements, the ow is pushed back by sending a backpressure message upstream to the source - no time and e ort was spent so far for reservations. As a result of backpressure, the incoming ow is either rerouted or rejected. Once the inelastic ow is established, it cannot be displaced by incoming inelastic ows because of a built in priority. The backpressure works only if the congested link is shared with proportional fairness among inelastic contenders in the same contention domain. To understand this concept, note that in the wired Internet, when a link becomes congested and the queue over ows, the packet drop rate of each ow is proportional to its rate, namely, a drop probability is uniform across ows. Without loss of generality, assume all ows are inelastic. If a new probing ow nds enough capacity on intermediate links and su ers no loss, it successfully completes the call set up and is promoted to established ow with higher drop threshold. If the new ow does not t in the bottleneck, i.e., it causes congestion, than it drops packets. Its drop probability is equal to that of the incumbent ows. By setting a lower drop probability threshold on new ow than on incumbent ows, the new ow is automatically discriminated and backpressured, leaving the incumbent ows undisturbed. We can easily discriminate and backpressure the new ow. A similar CAC approach was proposed several years ago for Internet VoIP streams [2]. In the Internet, where competing ows share a single common queue in the router, proportional fairness and more generally resource allocation are rather straightforward. In the wireless medium, there is no single common queue. In fact, there are several queues that are independently adjusting their MAC parameters (including retransmission rates) in case of loss. Thus, there is the risk of unfairness and channel capture by big ows and by ows with a relative interference graph advantage when the wireless medium becomes congested. Clearly, a distributed proportional fairness scheme must be developed to overcome the lack of a centralized control point. To this end, we have complemented the probing scheme with a distributed fairness scheme, Neighborhood Proportional Drop (N-PROD) which enforces uniform drop probabilities among ows competing in the same contention domain. Each node estimates own packet drop probability and propagates this information by piggybacking to neighbors. As mentioned earlier, in ProbeCast, the incoming ow has by design a lower drop probability threshold than the incumbent ows. If during probing, the new ow drop rate increases beyond the threshold, the ow is backpressured toward the source node and the ow is rerouted. If backpressure pushes the ow back to the source and all alternate routes are exhausted, ProbeCast reject the incoming ow. The problem of fair sharing among inelastic multicast ows and the concept of proportional fairness was introduced in a companion paper that appears in MSWIM 2008. This paper extends that work by designing a Call Admission Control scheme based on backpressure. The major contribution of ProbeCast is to enable CAC and fair allocation of inelastic ows in MANETs, for both unicast and multicast streams, without requiring prior resource reservation and thus overcoming the overhead limitations of traditional MANET reservation and allocation CAC schemes. The rest of the paper is organized as followed: Section II illustrates the related work; Section III describes the details of the ProbeCast; Section IV presents simulation results. Conclusion and future work will be on Section V. categories based on their resource measurement and reservation methods. 2.1 Bandwidth Estimation and Resource Reservation Ad hoc QoS Multicasting (AQM) [3, 4] achieves multicast QoS support by tracking available neighbor nodes resources. Nodes periodically broadcast a hello message including own bandwidth usage. Upon receiving the hello message, nodes record neighbor information in a neighborhood table which is used to calculate the total bandwidth allocation to existing multicast sessions. When starting a multicast session, a node oods an initiation packet. Intermediate nodes forward it on feasible links based on the neighborhood table. AQM hello messages introduce considerable overhead in a mobile network, interfering with QoS support. The Lantern tree (LTM) [5] relies on a multipath structure, called a lantern-path. LTM employs the lantern tree as a routing path in ad hoc multicast and it uses a CDMA-overTDMA model at the MAC layer to allow the superposition of many ows. LTM exploits CDMA orthogonal multiuser capability to allocate an extra ow in an already occupied network. QoS is guaranteed only to the extent that the load is kept in check (else losses escalate). Main implementation drawback is the need for a non standard CDMA-over-TDMA MAC with distributed time synchronization requirements. QoS Multicast Routing Protocol (QMR) [11] is an ondemand mesh based protocol that uses a forwarding mesh like On-Demand Multicast Routing Protocol (ODMRP) [8]. QMR de nes Forward Nodes (FNs) which establish a forwarding mesh and provide multiple paths. FNs reserve bandwidth for a multicast session if they can accept QoS route request (QREQ) from the source. Upon receiving data packets from multicast sessions WITHOUT reservations, they forward them only if shared bandwidth is available. To implement this hybrid scheme, nodes divide bandwidth into x reserved and shared bandwidth. The mesh structure can guarantee good delivery ratio via redundant forwarding. However, ood type redundancy may lead to congestion and excessive overhead e ecting QoS performance of reserved ows. 2.2 End-to-End Probing 2. RELATED WORK A number of ad hoc unicast QoS support protocols and algorithms have been proposed in the literature, e.g., INSIGNIA [7], SWAN [1], and FQMM [6]. However, relatively few MANET multicast QoS schemes have been introduced. We classify existing MANET multicast QoS schemes into 3 Multicast- Call Admission Protocol (M-CAMP) [10] is an end-to-end probing protocol in which a source, before transmitting the data stream oods probing packets to test bandwidth availability along the multicast tree. Only the receivers participate in the CAC decision by submitting an accept/refuse decision to the source based on the received quality. Similar to PCP [2] M-CAMP employs 3 priority levels among packets: real time, probe, best e ort. Level 2 probing packets do not a ect existing QoS ows. To cope with mobility, after topology changes, a new resource probing process is started to rebuilds the tree and the allocation. QAMNet [13] establishes a QoS aware mesh using JoinProbe and Probe-Response control packets. QAMNet exploits MAC layer feedback to estimate available bandwidth. Like QMR, a source oods a Join-Probe packet when it has packets to transmit. Intermediate nodes update a bandwidth eld in the Join-Probe packet to re ect minimum available bandwidth along the path. After collecting one or more Join-Probe packets, a receiver sends a Probe-Response to source if a feasible path is found. The main drawback of both end to end schemes is the inability to prevent unfairness and capture. In addition, QAMNet incurs the burden of local available bandwidth estimation. 2.3 Bandwidth Fair Sharing As mentioned earlier, Mar a et al. designed an algorithm, called FairCast [9]. The main focus of FairCast is fair sharing across multicast ows. FairCast does not exercise Call Admission Control, i.e., it cannot reject ows when congestion sets up. To cope with this situation, FairCast assumes that inelastic ows have su cient erasure coding redundancy so that they can tolerate even substantial losses. Alternatively, the inelastic ows can be adaptively rate adjusted. FairCast uses only local ow interactions and packet dropping to achieve fairness; no end-to-end feedback. FairCast rst introduced the concept of distributed proportional fairness in wireless channels. Flows locally interact, exchanging (piggybacked) information on packet loss rates and selectively drop packets to equalize their performance. FairCast di ers from ProbeCast in that it does not address Call Admission Control. Algorithm 1 Calculates The Node Drop Probability DEFINITIONS: For each ow r, DqNir is the number of dropped packets at the queue in a unit time, DlNir is the number of dropped packets on the link, and RNir is the number of received packets at node i prior to queue drop. Node monitors DqNir , DlNir , and RNir . Pir is packet drop probability of ow r at node i. P robi is the Node Drop r Probability of node i and T hri is the drop threshold of ow r. for each ow r rin i do r DqNi (t)+DlNi (t) r Di (t) = RN r (t)+DlN r (t) i i r Pir (t) = Pir (t 1) + (1 )Di (t) r r if Pi (t) > T hri then P robi = Pir (t) SetBackpressureFalg(r) else if Pir (t) > P robi or P robi is timeout then P robi = Pir (t) RecordtheUpdatedTime(P robi ) end if end for 3. PROBECAST In this section, we present a detailed description of ProbeCast. 3.1 Assumptions In ProbeCast, a key underlying assumption is that inelastic ows are protected by some form of end to end FEC (eg, erasure coding, fountain codes, raptor coded, etc). Namely, a sender adds redundancy to its stream, in the form of error correcting code which allows a receiver to detect and correct errors (within some limits) at the expense of some extra delay, without the need for retransmission. This is critical in MANET multicast sessions since conventional ACK and retransmission techniques between a sender and receivers may cause ACK/NACK implosion. We also assume that inelastic ows are classi ed into several priority levels. Each level is given a maximum tolerable loss rate which (in the erasure code implementation) corresponds to a packet drop threshold. The packet drop threshold is carried in the packet header. An intermediate node backpressures the ow when the packet drop rate exceeds the threshold. ProbeCast is independent of the underlying multicast routing protocols (in our simulation experiments we will use ODMRP). ProbeCast works equally well with unicast (a special class of multicast). It can coexist with lower priority best e ort tra cs (e.g., TCP), and will throttle best e ort tra cs to make room for inelastic, higher priority ows. single MANET node. When transmitting a data packet, an intermediate node updates the local sequence eld in the packet header. Upon receiving a packet, a down-stream node increments the number of received packets and monitors the local sequence number in the packet header. If there is a gap, a packet was lost. A node also tracks the number of packets it drops from its queue. It does not, however, attempt to monitor packet drops on outgoing links to neighbors. This count is the responsibility of the downstream neighbors, which eventually report the loss to upstream nodes. Every time unit, a node estimates its packet drop rate. Since the estimate uctuates, ProbeCast uses a weighted average to smoothen uctuations. Drop probability computation follows: DNir (t) = DqNir (t) + DlNir (t) r Di (t) = (1) (2) (3) RNir (t) DNir (t) + DlNir (t) r Pir (t) = Pir (t 1) + (1 )Di (t) where: t is the t-th time interval DNir is the total dropped packet rate at node i, ow r DqNir is the dropped packet rate at the queue at node i, ow r DlNir is the dropped packet rate on the incoming link to node i, ow r RNir is received packet rate at node prior to drop i, ow r r Di is the calculated packet drop rate at node i, ow r 3.2 Packet Drop Probability ProbeCast delivers drop probability information via piggybacking, in the packet header. To calculate the packet drop probability, each intermediate node keeps a time window based revolving count of received and lost packets. In addition to the sequence number stamped by the source and used to discard duplicates, each intermediate node keeps track of ows and assigns local sequence numbers to packets in each ow. Local ow bookkeeping is generally unacceptable in the wired Internet because of scalability considerations. In our case, scalability is not violated due to the rather limited number of simultaneous inelastic ows in a Pir is the packet drop probability at node i, ow r is the constant value, called step constant T hrr is the drop probability threshold for ow r Algorithm 1 summarizes the Node Drop Probability computation based on the above formulas. If a node relays multiple inelastic ows, each ow may have a di erent packet drop probability. To reduce overhead, instead of sending all drop probabilities, it su ces for a node to propagate just the highest value. For convenience, we call this value the Node Drop Probability hereafter. Upon hearing the neighbor Node Drop Probability, a node sets own Node Drop Probability by neighbor s value if neighbor s Node Drop Probability is higher than its own value. Node Drop Probability values are timed out and refreshed to account for lossy neighbors that move away. Algorithm 2 N-PROD Algorithm r DEFINITIONS: Di is packet drop rate of ow r at node i. r P robi is the Node Drop Probability of node i and T hri is r s the drop threshold of ow r. pkt and pkt are the packet of ow r and s, respectively. Node i receives pkts from Node j Node Drop Probability of i is P robi N umReceivedP kt = N umReceivedP kt + 1 if P robj > P robi and P robj and P robi are di erent ows then P robi = P robj RecordtheUpdatedTime(P robi ) end if Node i sends pktr Node Drop Probability of i is P robi and it is ow s while queue is not empty do if packet is pktr and P robi is not ow r then if P robi > unif ormRandom[0, 1] then PacketDrop(pktr ) end if end if pktr DropP rob = P robi PacketSend(pktr ) end while own Node Drop Probability by neighbor s value; otherwise, the node ignores it. To enforce drop probability, before forwarding a packet, the node generates a random number to compare it with the target Node Drop Probability. If the number is smaller than the Node Drop Probability, the packet is dropped from the queue; otherwise, it is forwarded. As a exception, heuristic the packet is not dropped if it belongs to the ow with highest Node Drop Probability, in order not to further hurt that ow. 3.4 Backpressure 3.3 Neighborhood Proportional Drop N-PROD allows inelastic ows to acquire resources in a fair and totally distributed manner without resource reservation. It enforces proportional drop rates among ows competing in the same contention domain. Note that proportional fairness is not generally desirable in elastic ows such as best e ort data sessions controlled by TCP. In fact, a popular TCP fairness scheme called NRED [14] enforces uniform drop probability so that all the TCP ows in the same contention domain achieve the same throughput. Proportional drop in N-PROD can enforce di erent throughputs for di erent inelastic ows with di erent nominal rates. For example, if ow A and B send 100Kbps and 60Kbps respectively and N-PROD control stabilizes at 20% drop probability, ows A and B throughputs stabilize at 80Kbps and 48Kbps respectively. In contrast, TCP fairness strives to equalize ows. In our implementation 3.2, each node reads Node Drop Probability values from overheard packets and adjusts its Node Drop Probability value. If Node Drop Probability of the neighbor node is higher than its value, a node replaces The ow packet drop threshold depends on tra c class, encoding rate and age of the ow. For example, assume three inelastic ows have 50%, 40% and 30% drop thresholds respectively. The rst ow is more loss tolerant than the others. It will be more di cult to reroute or reject it once it is established. By the same argument, a new entering ow typically has a lower drop threshold than existing ows and thus it is the rst to be rejected in case of overload. When the packet drop rate is over the threshold, the ow is backpressured towards its source. The backpressure mechanism uses piggybacking to reduce overhead. Upon getting a backpressure signal from a neighbor, the node checks if the neighbor is one of its downstream forwarders for that ow. If so, it will remove the downstream node from the list. It will then check the list to determine if there are any other downstream forwarders or local receivers for the ow in question. If there are none, the node will forward the backpressure signal to its upstream node. This way, all non productive branches of the multicast tree are pruned. If the backpressure signal reaches the source, the ow is rejected (i.e., there is no receiver ready for it). Alternatively, the source can attempt to construct a new multicast tree/mesh by searching for lightly loaded paths. Figure 1 illustrates an example of new ow rejection via backpressure. In gure 1 (A), two inelastic ows, Flow 1 and 2, are initially allocated to two non interfering paths. In gure 1 (B), a new ow, Flow 3 is injected by ProbeCast. It starts transmitting packets which interfere with Flow 1 and 2. Thus, Flow 3 shows low delivery ratio since must compete against the other two ows. Finally, Flow 3 drop rate goes over the drop threshold and in gure 1 (C) backpressure and rejection occurs. The other ows are restored to their original rates. 4. SIMULATIONS In this section, we validate N-PROD and ProbeCast using Qualnet v3.9.5 [12], a packet level network simulator. We implement N-PROD and ProbeCast in Qualnet and compare the performance of ProbeCast to pure ODMRP. 4.1 Simulation Setup We use 802.11b with 376m e ective reception range and 2Mbps channel capacity. The packet size is 512 bytes and the maximum queue size is 50Kbyte (about 100 packets). We use Qualnet default values for MAC and Physical layer con guration parameters. ProbeCast and N-PROD can run on any ad hoc multicast routing protocol. In our simulation, we chose ODMRP. In ODMRP, a source periodically oods a Join Query packet into the whole network. Upon receiving a non-duplicate Join Query packet, every node in the network stores the upstream node address for reverse path learning and rebroad- Flow 2 Flow 2 Flow 2 Backpressure Flow 3 Flow 3 Flow 1 Flow 1 Flow 1 (A) (B) (C) Figure 1: 3 ows in the simple topology. Lower graphs show packet delivery ratios, presented by percentage. (A) Two ows are present both with have high delivery ratios over 90%. (B) Flow 3 starts transmitting and other ows delivery ratios decrease because of channel contention. (C) Flow 3 packet drop rate is exceed the threshold and backpressure start. S3 Flow 3 F3 R3 casts it. When the Join Query reaches a multicast receiver, the receiver creates and broadcasts a Join Reply packet to its neighbors. This Join Reply packet is relayed all the way back to the source following the learned reverse path. Nodes on the reverse path become the forwarding group. Data is delivered along the mesh consisting of the forwarding group nodes. We use the ODMRP implementation included in the Qualnet package. The Join Query refresh interval is 3 seconds and the forwarder life time is 3 times the Join Query refresh interval. We use three metrics: Throughput is the total received data bits divided by the total simulation time; Packet Delivery Ratio is the fraction of received data; Number of Packet Sent is the aggregated number of packets sent by a source. All numbers are averaged over 100 simulation runs except for the Number of Packet Sent. S2 F2 R2 Flow 2 F1 R1 4.2 3 Unicast Inelastic Flows - No Backpressure Flow 1 S1 Figure 2: 3 parallel inelastic ows topology example. Intermediate nodes, F1, F2 and F3 are within radio range and they compete with each other. Sources, S1, S2, and S3 are outside of other s radio range. We rst tested N-PROD ability to enforce proportional unfairness. In the process, we also show the di erence between uniform and proportional fairness. The scenario is a unicast tra c scenario shown in gure 2. The 3 inelastic ows in gure 2 use disjoint paths, but they still interfere with each other. For each ow, the source and destination are located out of each others transmission range and communicate only with intermediate nodes, more precisely node F1, F2 and F3. The distances between node F1, F2 and F3 are 350m and thus they hear each other and compete for the medium. The ows are inelastic; the sources, S1, S2, and S3, send data at a constant, uniform rate = 500Kbps. Flow 1 starts transmitting data 1 second after simulation initialization. Flow 2 and 3 start data sending T =10 second and 20 second respectively. Because node F2 is located within F1 and F3 s transmission range, Flow 2 packets are at a disadvantage and are dropped at F2 at a higher rate than the other ows. As a result, only a few Flow 2 packets reach the destination. The result is shown in Figure 3. The 3 Even Parallel Flows Throughput Throughput (Kbps) Throughput(Kbps) 4 50 400 3 50 300 2 50 200 150 10 0 50 0 3 Even Parallel Flows Total Throughput 1000 800 600 400 200 0 ODMRP with no N-PROD Protocols N-PROD Flow 1 Flow 2 Flow 3 ODMRP with no NPROD N-PROD Protocols Figure 3: Throughput of 3 inelastic ows. Uniform nominal rate = 500Kbps. Figure 4: Total throughput in the network in 3 inelastic ows case. Data rate = 500Kbps. 3 Uneven Parallel Flows Total Throughput 1000 Throughput(Kbps) 3 Uneven Parallel Flows Throughput 700 600 500 400 300 200 100 0 ODMRP with no NPROD N-PROD Throughput (Kbps) Flow 1 Flow 2 Flow 3 800 600 400 200 0 ODMRP with no N-PROD N-PROD Protocols Protocols Figure 5: Throughput of uneven 3 inelastic ows case. Flow 1, 2, and 3 are 800Kbps, 400Kbps, and 200Kbps respectively. Figure 6: Total throughput in the network of uneven 3 inelastic ows case. Flow 1, 2, and 3 are 800Kbps, 400Kbps, and 200Kbps respectively. application of N-PROD restores fairness as shown in Figure 3. This result is very similar to the result reported in paper [14]. This is not surprising since with uniform inelastic rates, N-PROD is equivalent to NRED. Figur 4 shows total throughput respectively. As already noted in [14], fairness comes at the cost of degraded total throughput. In the next simulation experiment, we use the same layout but now the inelastic ows send data packets at di erent rates. Namely, Flow 1 = 800Kbps and Flow 2 = 400Kbps and Flow 3 = 200Kbps. Like in the previous experiment, Flow 1 starts transmitting data at T =1 second. Flow 2 and 3 start transmissions at T = 10 second and T=20 second respectively. Without N-PROD, Flow 1 and 3 capture the channel and Flow 2 is starved. Flow 3 delivers more than 90% of the sent packets while only 10% of packets are reached at the destination in Flow 2. Consequently, Flow 3 throughput, about 180Kbps, is higher than Flow 2 throughput, 55Kbps, while S2 sends packets with higher rate than S3. With N-PROD each ow drops packets proportionally to its demand. Thus, drop probabilities are uniform and achieved throughputs are staggered as the demands (in the ratios 8:4:2) as shown in gure 5. This result is clearly di erent from what could be achieved with NRED. The gure 6 represents total throughput in the network without and with N-PROD. In spite of the fact that individual throughputs are now proportional to demands, it appears that the total throughputs are rather insensitive to the actual distribution of demands. 4.3 More Realistic Scenario In the previous experiment, the topology was simple and was speci cally chosen to illustrate the di erence between uniform and proportional dropping and the importance of the latter in the support of non uniform inelastic ow. Moreover, the scenario was unicast, and no backpressure was enacted. In this section, we report on multicast experiments with ProbeCast, this time combining N-PROD, backpressure and CAC. Figure 7 is the topology example we used. 30 nodes are randomly distributed in a 1000m by 1000m area; 3 multicast sessions are established. Each multicast session has one source and 3 receivers and no common node belongs to two sessions. However, interference occurs at intermediate forwarding nodes in the eld. Inelastic data rates are uniform for simplicity, namely, 500Kbps for each ow. These sessions can tolerate up to 50% of packet loss (that is drop threshold is 50%). Multicast session 1 starts transmitting at T=1 second and session 2 and 3 start at T =10 and 20 second, respectively. In gure 8, we report the result for an experiment with only two multicast sessions (session 1 and 2). We performed several simulation run changing seed numbers. In almost all the runs, both sessions survive and manage to transmit their full rates. However, when all three sessions are injected, the results reported in gure 9 shows that one of three sessions is consistently rejected. At the beginning the sessions try to balance drop rates and partially succeed. However, as time progresses, this balance collapses. One session starts dropping packets in bursts and packet drop rate suddenly skyrockets, exceeding the thresh- Multicast Session 3 4000 3500 3000 2500 2000 1500 1000 500 0 40 SESSION 1 SESSION 1 Number of Packets Send Number fo Packets SESSION 2 Number of Packets Send 4000 3500 3000 2500 2000 1500 1000 500 0 40 80 0 20 0 12 0 16 0 24 0 SESSION 2 SESSION 3 Multicast Session 2 Multicast Session 1 80 0 20 0 24 0 12 0 16 0 28 0 Number of Packets T ime (s) T ime (s) Figure 7: 3 multicast sessions in the Figure 8: 2 multicast sessions can Figure 9: 3 multicast sessions are 1000m by 1000m area. Each session be simultaneously supported. has 1 source and 3 members. present. Session 2 is rejected. old and triggering backpressure on the newcomer (with lower drop threshold). Flow 2 S2 TCP Flow F2 4.4 Inelastic Flow vs. Elastic Flow The next experiment is designed to show that probing allows an inelastic ow to preempt an elastic ow (say TCP) by properly exercising the proportional drop threshold. Figure 10 represents a very simple network topology where an elastic video stream ow coexists with an inelastic TCP ow. The TCP sender, S2, starts at t =1 second; the inelastic sender, S1, starts at t=10 second. Video stream rate is 500Kbps. S1, S2, F1, and F2 are all within radio sensing range so that they interfere with each other but cannot decode each other transmissions. R1 and R2, however, are assumed far apart to reduce hidden terminal collisions (i.e., R1 is not interfered by F2 and vice versa). Note that the typical reservation based CAC scheme does not work in this situation. When S1 monitors the channel for available bandwidth, it nds none. In fact, it cannot tell that the interferer is a lower priority best e ort ow since distance exceeds reception range. On the other hand, the TCP ow (due to its greedy nature) completely lls the channel. Therefore, the inelastic ow is rejected. In contrast, ProbeCast lets the inelastic ow in, causing an increase in packet loss that in turn forces TCP to back o and leave enough room for the inelastic ow to achieve full rate. It is interesting to note that the TCP source will backo even if it does not hear the inelastic N-PROD signals (i.e., current drop rate). The interference and subsequent loss rate will su ce to slow down TCP. Figure 11 shows the number of sent packet at ow sources and gure 12 illustrates receiver s packet drop rate of the inelastic ow, R1. In gure 11, the middle line (triangle markers) represents the number of packet sent by TCP when it is alone. Comparing the two TCP lines (before and after) we notice a 3:1 degradation in TCP rate. Figure 12 shows an inelastic ow packet drop rate in the order of 5%. This loss is easily sustained as it is below the threshold and it is recovered by end to end erasure coding. The inelastic loss is in necessary to keep TCP at bay. The TCP ow in fact experiences a comparable loss rate on the shared channel. 300 m S1 F1 Flow 1 Inelastic Flow R1 Figure 10: Two ows. Flow 1 is an inelastic ow and Flow 2 is an elastic ow. which supports e cient call admission control and QoS in the MANET without requiring bandwidth estimation and reservations. ProbeCast uses probe and backpressure mechanisms to accept feasible ows and reject the unfeasible ones. For backpressure to work, a congested link must be fairly (more precisely, proportionally) shared among inelastic contenders. We apply a distributed fairness scheme, N-PROD, inspired to an earlier TCP fairness scheme (NRED) and to the Distributed Gentlemen s Agreement proposed in FairCast, to proportionally share bandwidth among ows. Major contributions of ProbeCast are: the robust and e cient CAC mechanism based on probing; the ability to handle both uni and multicast inelastic CAC, and; the ability to handle both inelastic and elastic ows at the same time. Simulation results show con rm our claims. Future work will examine the use of QoS multicast tree construction heuristics to facilitate rerouting in case of CAC blocking. Acknowledgment This research is supported through participation in the International Technology Alliance sponsored by the U.S. Army Research Laboratory and the U.K. Ministry of Defense under Agreement Number W911NF-06-3-0001, and; by ARMY MURI under funding W911NF0510246. 5. CONCLUSIONS AND FUTURE WORKS Inelastic multicast bandwidth allocation and CAC in MANETs is extremely challenging because of dynamic tra c and route changes and unreliable estimation of available bandwidth. In this paper, we have proposed a scheme called ProbeCast 6. REFERENCES [1] G.-S. Ahn, A. T. Campbell, A. Veres, and L.-H. Sun. Supporting service di erentiation for real-time and best-e ort tra c in stateless wireless ad hoc networks 28 0 R2 The Number of Packet Sent 40000 35000 30000 25000 20000 15000 10000 5000 0 30 60 90 0 12 0 27 0 15 0 18 0 21 0 24 0 TCP Inelastic TCP only Packet Drop Rate 20 Packet Drop Ratio (%) 15 10 5 0 30 60 90 0 24 0 12 0 21 0 The Number of Packets 30 0 T ime (s) Figure 11: The number of packet sent by the inelastic and the elastic sources. Figure 12: The packet drop ratio at the inelastic ow receiver in the two inelastic and elastic ow case. [2] [3] [4] [5] [6] [7] (swan). IEEE Transactions on Mobile Computing, 1(3):192 207, 2002. F. Borgonovo, A. Capone, L. Fratta, M. Marchese, and C. Petrioli. Pcp: A bandwidth guaranteed transport service for ip networks. In IEEE International Conference on Communications 1999 (ICC 99), volume 1, pages 671 675, 1999. K. B r and C. Ersoy. Multicast routing for ad hoc u networks with a multiclass scheme for quality of service. In 19th International Symposium on Computer and Information Sciences (ISCIS), pages 187 197, 2004. K. B r and C. Ersoy. Multicast routing for ad hoc u networks with a quality of service scheme for session e ciency. In 15th IEEE International Symposium on Personal, Indoor and Mobile Radio Communication (PIMRC 2004), pages 1000 1004, 2004. Y.-S. CHEN and Y.-W. KO. A lantern-tree-based qos on-demand multicast protocol for a wireless mobile ad hoc network(network). IEICE transactions on communications, 87(3):717 726, 20040301. H.Xiao, K.chua, W. Seah, and A. Lo. A exible quality of service model for mobile ad-hoc networks. In Proceedings of Vehicular Technology Conference (VTC), pages 445 449, 2000. S.-B. Lee, G.-S. Ahn, X. Zhang, and A. T. Capbell. Insignia: an ip-based quality of service framework for mobile ad hoc networks. Journal of Parallel and Distributed Computing, 60(4):374 406, 2000. [8] S. J. Lee, W. Su, and M. Gerla. On-demand multicast routing protocol in multihop wireless mobile networks. Mobile Networks and Applications, 7(6):441 453, 2002. [9] G. Mar a, P. Lutterotti, S. J. Eidenbenz, G. Pau, and M. Gerla. Faircast: Fair multi-media straming in ad hoc networks through local congestion control. In 11th ACM International Conference on Modeling, Analysis and Simulation of Wireless and Mobile System, 2008. [10] E. Pagani and G. P. Rossi. A framework for the admission control of qos multicast tra c in mobile ad hoc networks. In Proceedings of the 4th ACM international workshop on Wireless mobile multimedia(WOWMOM 01), pages 2 11, New York, NY, USA, 2001. ACM. [11] M. Saghir, T. C. Wan, and R. Budiarto. Load balancing qos multicast routing protocol in mobile ad hoc networks. In Proceedings of Asian Internet Engineering Conference (AINTEC 2005), pages 83 97, 2005. [12] Scalable Networs Inc. QualNet. http://www.scalble-networks.com. [13] H. Tebbe, A. J. Kassler, and P. M. Ruiz. Qos-aware mesh construction to enhance multicast routing in mobile ad hoc networks. In Proceedings of the rst international conference on Integrated internet ad hoc and sensor networks (InterSense 06), page 17, New York, NY, USA, 2006. ACM. [14] K. Xu, M. Gerla, L. Qi, and Y. Shu. Tcp unfairness in ad hoc wireless networks and a neighborhood red solution. Wirel. Netw., 11(4):383 399, 2005. 15 0 T ime (s) 18 0 27 0 30 0

Find millions of documents here - Study Guides, Homework Solutions, Papers, Exam Answer Keys and more. Course Hero has millions of course related materials that will enable you to learn better, faster and get an A in all your courses.
Below is a small sample set of documents: