Course Hero - We put you ahead of the curve!
You have requested the below document.
- Title: faircast
- Type: Notes
- School: UCLA
- Course: CS 468
- Term: Fall
FairCast: 1 Fair Multi-Media Streaming in Ad Hoc Networks through Local Congestion Control Gustavo Mar a , Paolo Lutterotti , Stephan J. Eidenbenz , Giovanni Pau , Mario Gerla Computer Science Department - University of California, Los Angeles, CA 90095, US e-mail: {gmar a|gpau|gerla}@cs.ucla.edu CCN5 - Los Alamos National Laboratories - Los Alamos, NM, US e-mail: eidenben@lanl.gov Istituto Superiore Mario Boella, Torino, 10138, Italy e-mail: lutterotti@ismb.it Abstract Multicast streaming is gaining increasing importance in wireless ad hoc networks, in part because ad hoc scenarios often include team activities and the requirement for distribution of audio, video and situation awareness to the members. At the network level, techniques for routing the multimedia streams are quite mature. Much more challenging is the allocation of resources, the fair sharing among streams and the control of congestion. While in rate adaptive UNICAST streams congestion control and fair sharing are accomplished with end-to-end feedback techniques inspired to TCP, the feedback does not scale well in MULTICAST. In fact, it leads to the well knows ACK/NAK implosion problem and unfair penalties for heterogeneous receivers. These limitations can be overcome using backpressure from congestion points to the sources - but this approach suffers of latency and cannot rapidly adjust to changes in traf c. Another solution is multilayer adaptive coding. Namely, the encoding adaptation is done locally by dropping layers. It does not require end-to-end feedback nor changes in input rates. Multi-resolution codes are now becoming attractive due to the progress in technology; we expect these to become the prevalent techniques in large scale media distribution. One issue, however, that still remains to be local resolved is the fair sharing among competing multicast streams. In this paper we address the congestion control AND fair sharing in a multilayer multicast scenario. We show that lack of proper fairness provisions in the local adjustments can lead to serious capture situations, especially in heterogeneous traf c mixes (e.g. voice and video). We Reference Author: Gustavo Mar a, Computer Science Department, University of California Los Angeles, CA 90095, email:gmar a@cs.ucla.edu This work is partially supported by the Italian Ministry for Research via the ICTP/E-Grid Initiative and the Interlink Initiative, the National Science Foundation through grants Swarms and Whynet, and the UC-Micro Grant MICRO 04-05 private sponsor STMicroelectronics. Any opinions, ndings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily re ect the views of the National Science Foundation. then propose a FAIR local adjustment that targets a fair dropping of packets in each interference domain. We show that the scheme can be interpreted as a distributed implementation of a utility function minimization, where the utility is the packet loss subject to fairness bounds across ows. This formulation guarantees stability and convergence of the distributed algorithm. The main contributions of this paper are the low overhead design of the local fairness enforcement algorithm, the utility function framework and the demonstration of convergence via simulation in representative scenarios. I. INTRODUCTION We here propose a novel approach to multicast congestion control and fairness, based on local ow interaction. Differently from traditional models, where a ow is controlled by end-to-end feedback signals (e.g. packet loss and delay), we build an in-network, distributed ow interaction mechanism where no explicit control packets are delivered from destinations to sources. Competing ows locally interact, exchanging information on loss rates and adapting to each other s performance. This approach is appropriate for ad hoc networks where the cost of sending end-to-end control packets is expensive. It is also more responsive to highly dynamic traf c and topology changes. In this study we focus on fairness. In particular we examine the fairness problem for real-time multimedia multicast ows. The fairness problem was thoroughly studied in the case of unicast best effort (e.g. TCP) and real time sources for wireless ad hoc networks. However, to our knowledge, little progress was done so far in ef cient and fair multimedia multicast. Ad hoc networks have been proposed for a variety of applications from battle eld missions to disaster recovery and vehicular communications. In any of these cases, streaming-based applications play an important role. In a disaster recovery application, for example, team leaders 2 Delivery Ration at 50 Pkts/s 1.0 0.9 Stream 0 Stream 1 Stream 2 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Fixed Routing Fig. 1. Unfairness between three ows. will need to multicast orders by voice to coordinate rescue teams, while the members of a rescue team will want to multicast a situation update of the disaster scene. Voice over IP (VoIP) and peer-to-peer streaming will also become popular in VANETs and metropolitan mesh networks. In brief, we can expect that the multi-hop wireless networks of the future, both xed and mobile, will carry a great deal of voice/video streaming. For an ef cient utilization of such networks, in terms of the number of streaming ows they are able to serve, rather than the aggregate throughput they deliver, a fair sharing of resources between ows is required. A simple case of unfairness due to spatial contention is illustrated in Fig. 1. We plot the delivery ratio for three ows, on a 12 x 3 grid topology. Flow i originates from node (1, i) and terminates at node (12, i), so that each packet travels at least 11 hops. The delivery ratio of three competing CBR ows is plotted for three types of routing. The ideal routing solution, represented by parallel xed routes, penalizes the goodput of the central stream, although improving overall utilization, as there is no routing overhead and inef ciency. If these were video streams capable to tolerate up to 30% loss thanks to a robust coding technique, 90% delivery ratio for ows 1 and 3 and 45% delivery ratio for ow 2 signi es that only ows 1 and 3 will be adequately served. Now, if ows 1 and 3 were reduced by 20%, say, thus freeing enough capacity for ow 2 to increase by 25% its delivery ratio, we would on one hand loose in aggregate utilization while on the other increase ow utilization by 50% (i.e. three ows instead of two would be adequately served). The unfairness problem is even more severe when considering multiple multicast ows. In such case a multicast stream can choke others. To underline the problem we show a case of unfairness between two multicast groups routed by the On Demand Multicast Routing Protocol (ODMRP) [13], a multicast protocol which forwards packets through a mesh structure, thus providing multiple routes from a source to each destination. ODMRP was chosen as it is popular and readily available in simulation platforms - the same behavior is expected of any ad hoc multicast scheme. We examine how two multicast groups coexist by rst running one multicast group by itself and then adding the second. We here show the results for the two scenarios. The bit rates of the two sources of the two multicast groups are chosen to emulate a low quality video multicast and an audio multicast. The MAC layer we use throughout the paper is 802.11b, with rate xed at 1Mbps, thus disabling auto rate fallback. In the rst test only the video multicast group is active. In the second experiment both, video and audio, are active. The streams are supposed to be layer encoded. For simplicity, in the following we will assume a very ne layer grain (in the limit, an in nite number of layers, each with in nitesimal rate), such that we do not have to be concerned about the discrete nature of the adaptive multilayer control. Moreover, we assume that within each ow the packets are dropped by layer priority. The network is composed of thirty static nodes, the terrain is a 1000 x 1000 meters square and nodes are displayed using Qualnet s uniform placement model (i.e. the terrain is divided in as many cells as the nodes are and, within each cell, a node is randomly placed). Nodes are approximately placed as in Fig. 2. Both multicast groups have a single source, while nine receiving nodes are in the rst group and three in the second. Moreover, no receiver or source in one group is a receiver or a source in the other group. Simulations last 560 second. The source of the rst multicast group sends packets at a rate of 50pkts/sec or 200kbps, while the source of the second multicast group sends packets at a rate of 10pkts/sec or 40kbps. The gure of merit is the delivery ratio that is directly related to the received video quality. Results are averaged over the 200 simulations that are run per each set and show the overall delivery ratio per each receiver. Results for the rst set of simulations, where only multicast group one is active, are shown in Fig. 3. Results for the second set of simulations are shown in Fig. 4. The delivery ratio of multicast group one doesn t signi cantly change between experiments. Multicast group two is heavily penalized in the second set of simulations. The members of multicast group two, mainly due to the contention with multicast group one, loose most of the packets along the way. At this point the question is whether a more fair distribution 3 ODMRP Average Delivery Ratio 1.0 0.9 0.8 0.7 Receiver 1 Receiver 2 Receiver 3 Receiver 4 Receiver 5 Receiver 6 Receiver 7 Receiver 8 Receiver 9 0.6 0.5 0.4 0.3 Fig. 2. Multicast example topology. 0.2 0.1 0 ODMRP 200Kbps ODMRP 40Kbps ODMRP Average Delivery Ratio 1.0 0.9 0.8 0.7 Receiver 1 Receiver 2 Receiver 3 Receiver 4 Receiver 5 Receiver 6 Receiver 7 Receiver 8 Receiver 9 Fig. 4. Two multicast streams, a video (i.e. 200kbps) multicast stream and an audio (i.e. 40kbps) multicast stream are sent from two different sources. We can observe an inter-multicast ows unfairness. 0.6 0.5 0.4 number of interfering neighbors, will receive a different share of resources. Flows, with no nodes in common along their paths, may still be competing for resources due to spatial contention. This leads some ows to capture most of the channel s bandwidth and others to fall into starvation. A number of solutions have been proposed to balance this effect for TCP sources [5], Fig. 3. Single video (i.e. 200kbps) multicast stream case. All receivers [6]. Authors in [5] extend the Random Early Drop receive approximately the same delivery ratio, no relevant unfairness (RED) [7] concept to ad hoc networks, by observing is observed. that the algorithm should be enforced on the distributed neighborhood queue rather than on a single node s of resources could be guaranteed. The answer is probably queue. FairCast is a mechanism which balances resourceaf rmative, since we see that between experiment one sharing among ows. In this spirit, it is similar to the and experiment two multicast group one s delivery ratio approach outlined in the NRED algorithm (though in our is almost unaffected. Almost all members of multicast case we deal with drop rates rather than queue lengths). group one suffer at most 20% loss. If multicast group A new TCP source algorithm, designed to fairly work one receivers were willing to take a 30% rate reduction, on wireless multi-hop networks is proposed in [6]. there would likely be enough space for multicast group A second stream of work looks at fairness between two as well, thus achieving a more fair access to the neighboring nodes [8] [12]. The main approach in the wireless resource. proposed solutions is to design new MAC layer protocols The problem of ensuring fairness has been long stud- or to modify the existing de-facto standard, 802.11a/b/g, ied in wired networks, especially in the case of elastic by modifying the scheduling and/or the backoff algoTCP sources. TCP strives to maximize the sum of source rithm. utility functions and the fairness characteristics of a ow can be directly drawn from the utility function it impleThe main contributions of this work are: (a) to enforce ments. These results are not easily mapped to contention- overall network fairness on adaptive rate multimedia based wireless networks, due to a major complexity of ows by implementing a local, distributed algorithm, the network model. Unfairness between competing TCP and; (b) to introduce a new feedback paradigm, based ows is aggravated, in multi-hop wireless networks, by on ow interaction, which is meant to add onto and not the channel capture effect. Nodes, with a different to substitute network feedback as packet loss and delay. 0.3 0.2 0.1 0 ODMRP 200Kbps 4 II. ALGORITHM The main assumptions in designing the algorithm are that ows compete for network resources and that ows can tolerate some degree of loss. The loss that a ow can tolerate depends both on the application and the encoding overlaid on it (e.g. in our case the number of layers required for acceptable reception). This loss rate is upper-bounded by a threshold that corresponds to the minimum acceptable quality. In case a ow experiences a loss rate that exceeds the desired threshold because of other ows aggressiveness, it promptly reacts against the competing ows requesting them to drop at a certain rate. This is the Distributed Gentlemen Agreement (DGA) implemented in FairCast. This means that if two ows compete and achieve different delivery ratios, the most penalized ow complains with the other and they both agree to a local fair share. To relate this to the traditional wired network congestion control framework, the shared wireless channel is here interpreted as a droptail router; ow reaction may be interpreted as the price of exceeding the capacity of such router (i.e. packet loss). We may nd the following advantages in such approach: (a) no bandwidth is wasted due to end-to-end feedback; (b) ows are faster in adapting to highly dynamic traf c changes (i.e. ows adapt to congestion and to traf c/topology changes on the y, avoiding any end-to-end delay between source and receiver); and (c) it is easy to implement (i.e. no modi cations are required below the network layer). We also nd an equal number of disadvantages: (a) local decisions on limited information imply this is a suboptimal solution; (b) a packet which is voluntarily dropped at some node to be nice to a competing ow wastes the bandwidth used on the path traveled so far; and (c) some overhead is added to data packets (FairCast ows locally exchange congestion signals with data packets, but this overhead may also include, depending on the application, extra coding to make the ow robust up to a desired loss threshold). We divide the algorithm implementation in two successive steps. The rst step assumes each node knows the average packet loss rate threshold per each ow. By knowing this information each node will try to keep the packet loss rate below this value and will not react if this is, on average, bounded by this value. The second step is to understand how to set the threshold. We will better understand how these algorithms are derived in the Sections III and IV, where we derive the DGA algorithm and discuss the convergence properties of the threshold algorithm. Fig. 5. Topology for the asymmetric unicast evaluation of FairCast. A 5-hop ow competes with a 1-hop ow. Delivery Ration at 90 Pkt/s 1.0 Stream 0 Stream 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 FairCast Fixed Routing Fig. 6. Unfairness between the two ows in Fig. 5. A. Distributed Gentlemen Agreement With FairCast ows locally agree on how to share bandwidth by exchanging packet loss rate information. In each interval a ow measures the amount of packets it received and the amount of packets it was able to send. This way a ow, on a per node basis, is able to determine its local packet drop rate. If the local packet drop rate, say, exceeds the threshold of ow r on node n for a certain time, ow r will ask the ows that have an impact on its own performance, in particular, the ows traveling with it and the ows traversing the neighbors of node n to lower their rates. As a simple illustration of this concept we consider the topology shown in Fig. 5. We use unicast ows for simplicity. There is a long and a short ow. Without fairness control, the short stream (stream #1) wins. FairCast forces the ows to share the channel fairly as shown in Fig. 6. B. Algorithms In each time interval t, which in the following formula is normalized to 1 for simplicity, each node calculates, per each ow, the following value: r + r,n (t+1) = [ r,n (t)+ (pr loss,n (t+1) thrloss,n (t+1))] (1) 5 where: r,n (t) is the drop probability set by ow r, at node n, for its competing ows , a constant, is the step-size of the gradient algorithm pr loss,n (t) ow r s packet loss rate r thrloss,n (t) the threshold above which ow r reacts to packet loss Flow r may exceed the maximum tolerated packet loss r rate thrloss,n (t), which we assume to be known at each node, per each ow. Periodically ow r calculates , a measure of how much ow r is loosing with respect to r threshold thrloss,n (t), implementing Algorithm 1: Algorithm 1 Lambda Algorithm Ensure: Each ow r, within node n, reacts to packet loss by computing r,n r.d[k] is number of packets which node n failed to forward in the k-th interval. This value includes both the packets that have not been acknowledged and packets that have been dropped by FairCast. r.routed[k] is the number of packets that node n received to forward in the k-th interval. is the exponential lter constant. is the gradient-algorithm s step-size. for each ow r in n do if r.d[k] > 0 then d rate[k] = r.d[k]/r.routed[k] r.d rate[k] = (1 )d rate[k] + r.drop rate[k 1] r,n [k] = r,n [k 1]+ (r.d rate[k] r.thr[k]) end if end for Once r,n is computed, this is then inserted in any outgoing packet of ow r. We avoid the overhead of sending a new control packet at the price of wasting some space on each data packet. What then happens is that some node m, interfering with node n, receives this packet. If the lambda node m receives is greater than the last received lambda or if the last received lambda expired, node m saves this value. It then runs the algorithm we show in Algorithm 2 every time it forwards a packet. We can summarize what happens with Algorithm 2 as follows: before sending a packet of a ow a node compares the local lambda of this ow to both the ow s threshold and the last valid lambda received at this node. If the local ow s lambda is smaller than both these values, it means the that ow may tolerate losing some packets. In such case, the node will lower the ow s Algorithm 2 Receive Lambda Algorithm Ensure: Node m, a neighbor of node n, receives r,n . Node m reacts with a local drop, if certain conditions are met. for next packet that is going to be sent do next packet to be sent is pkts of ow s r,n = Last received lambda if ( s,n < r,n )&&( s,n < s.thr)&&(r! = s) then prob drop = unif orm random[0, 1] if prob drop > 1 r,n then dropped packets = dropped packets + 1 DropP acket(pkts ) else if r,n > max local lambda then set forward lambda = r,n else set forward lambda = max local lambda end if end if end if end for access rate to the wireless medium (thus increasing its own loss rate), leaving some space for the competing node s ow that is suffering from contention. If the condition is not met with any ow, it is not possible to redistribute ow loss by penalizing a local ow and the originator of contention (to be reduced) should be found elsewhere. For this reason, the node still checks if the received lambda is higher than the locally calculated lambda. If this condition is met, the node will forward the received lambda instead of its own. In this way, a lambda can be propagated two hops from each node and it is likely that the congestion origin can be reached. III. GENTLEMEN AGREEMENT MODEL We here de ne an optimization model that will lead us to de ne the FairCast algorithm and protocol. The result of this section will be the following result: we will derive ow interaction rules and selective drops from a utility function optimization which will lead to an overall minimum packet loss and a bounded maximum packet loss per ow on each hop. We will interpret the resulting algorithms in the wireless ad hoc domain, showing how ows can adjust their rates to mitigate the unfairness due to spatial contention between nodes. In the following for simplicity of notion we formulate the problem considering a single multicast communication. Multiple unicast and multiple multicast models are straightforward extensions or reductions of this model. 6 From now on we will use the following notation: L is the set of links of the network, N is the set of nodes and Ri is the set of receivers for source i. pr loss is a vector of |L| rows, for receiver r. Row pr loss,l represents the average fraction of packets lost, for receiver r, on link l. r thrloss is a vector of |L| rows, for receiver r. Row l represents the maximum fraction of packets lost that receiver r, that can be tolerated on link l before complaining . r pr loss,0 and x0 are scalars that both depend from the data rate and the packet size of ow r. xr is a vector of |L| rows, for receiver r. Row xr l represents the average rate of ow r on link l. xsource is a scalar. This is the maximum rate at the source. r 0 is a constant. This is a constant per each ow and relates the average loss rate to the average rate received by a ow. A. Optimization Problem We de ne our problem as follows: min r 1T pr loss (2) subject to: r pr thrloss , r Ri loss (3) consistent in the measure that we simulate the network using Qualnet and feed the observed x and pc vectors loss into our optimization problem. We are then here de ning a Simulation-Optimization problem, a generalization of a Deterministic Optimization problem, where one or more of the constraints is observable through a stochastic simulation. The objective, expressed by eq. (2), is to minimize the fraction of packets lost in the network. Constraint (3) xes a packet loss upper bound at links. Constraint (4) linearly models the relation between packet loss and rate at each link, approximating it in a neighborhood of steady state. Eqs. (5) and (6) are feasibility constraints, where eq. (6) limits the total rate at receivers to be bound by the rate at source by the number of receivers. The rationale behind eq. (3) is that in a resource constrained environment such as an ad hoc network an information loss decision should be distributed. An endto-end formulation would bring to an optimal solution of the problem, but would also introduce overhead, resource consumption and feedback delay. This would lead to an infeasible practical resolution of the problem. We assume eq. (4) to hold in an interval of the average bandwidth required to send a ow at each link. Bandwidth may heavily oscillate in an ad hoc network, we will therefore see how this assumption holds from our simulation results. In general, there may be heavy uctuations around the average bandwidth in contention based wireless networks. We simplify the problem by merging (4) and (6) and relaxing the constraints. The new set of constraints is: r pr thrloss , r Ri loss r rr r pr loss,l ploss,0 0 (xl x0 ), l L, r Ri (4) (7) xr 0, r Ri xr |Ri |xsource 1 (5) (6) r 1r r pr 0 xr 1) |Ri |xsource 1 0 loss x (p 0 loss,0 0 pr pr loss loss,0 (8) (9) r where vectors pr = pr,c + pr,d , pr,c represents loss loss loss loss the fraction of packets lost due to contention, pr,d loss represents the fraction of packets lost due to FairCast drop decisions. Clearly, pr,d are the variables in the loss optimization problem. Please notice one important fact, an equation of type xr = F (x, p) is missing. Such l equation models how a rate of ow r over link l depends from the rates and loss rates of other ows on competing links. The de nition of such equation is out of the scope of this work, such equation depends on a number of factors and a closed form for such equation is still an open research problem. We will here assume x and pc loss are observable at each link. This assumption is We now decouple the problem in terms of ows, deriving |Ri | problems. Following the approaches discussed in [15], by relaxation, we de ne one sub-problem per each receiver and a single master problem to coordinate them. The |Ri | sub-problems are: minpr 1T pr + T loss loss 1r r pr 0 xr 1) (10) 0 loss r (p 0 loss,0 subject to: r 0 pr thrloss loss (11) 7 The master problem, which updates the dual variable in the sub-problems, is: max r fr ( ) |Ri |xsource T 1 (12) wireless spatial contention to be the dominant effect, so that packet loss is mainly caused by interference between ows and r is the ow reaction which leads ows to well behave . IV. THRESHOLD ADAPTATION MODEL subject to: 0 (13) A. Threshold Adaptation A problem we have not yet solved until now is how to choose the threshold values. We have shown that by implementing the algorithm derived in eq. (17) we are able to ensure fairness between competing ows, but threshold values were statically set to achieve the desired results. The threshold should instead be automatically set, dynamically adapting to traf c conditions. In fact, with adaptive layers, the threshold has a very precise meaning: the quality degradation a user is willing to tolerate - or, the min quality - minus a proper margin. Setting the threshold to minimum quality level, with zero margins, would guarantee that each ow gets at least the minimum acceptable level (i.e. we nd a feasible solution). For the solution to be optimally fair, however, the threshold must be as tight as possible. We here design a heuristic algorithm to solve this problem. The rationale behind the algorithm is as follows. The wireless channel, the clique on which ows contend, is the bottleneck for competing ows, just as a bottleneck router in the Internet. An Internet router, when the incoming ow exceeds its bandwidth, drops packets, thus sending a TCP source a congestion signal. In a similar way a ow that implements FairCast, say ow A, sends a congestion signal to other ows that share the same wireless channel when ow A is suffering from contention. Other ows compare the congestion measure they receive from ow A (i.e. ) and decide whether ow A is in a better or worst situation and whether they should drop packets. In the wired case a TCP ow, receiving a congestion noti cation (i.e. 3 DUPACKs), lowers the slow start threshold to mitigate network congestion. A FairCast ow should then, in a similar manner, increase its threshold to be more tolerant to loss. Here follows the algorithm: min{thrr (t) + a, 1}, if c (t) > r (t) max{thrr (t) b, 0}, otherwise (18) where a and b are xed parameters we will shortly investigate, c (t) the highest of the lambdas received from competing nodes at ow r and r (t) the lambda computed by ow r. In other words, if nobody complains, ow r keeps tightening up its threshold. We can thrr (t+1) = The dual variable, dimensionally [1/bps], is the link congestion signal which binds ow link rates to not exceed the available rate. A higher increases the fractional packet loss of one or more ows (i.e. we can see it as the equivalent of packet loss in TCP). We are now interested to build upon the sub-problems, represented by eqs. (10) and (11). We rst x l on each link and then write the Lagrangian of eqs. (10) (11): L(pr , s ) = 1T pr + loss loss T1 r + r (pr pr 0 xr ) + 0 loss 0 loss,0 + (pr )T pr loss loss r + T (pr thrloss ) + loss (14) the new term, (pr )T pr , is a regularization term loss loss which for 0 returns the Lagrangian to the original and the original solution is preserved. By adding this term we obtain feasible primal solution to the original problem. The dual of eqs. (10) (11) is (here is a constant, not a variable): max L(pr , r ) loss (15) The resulting gradient algorithm to solve the dual problem in (15) is: pr (t) = loss,l 1 l [ r 1 r,l ]+ 2 0 (16) r + r,l (t + 1) = [ r,l (t) + (pr loss,l thrloss,l )] (17) where is a positive step. here represents the link reaction to congestion, while r a node s s ow reaction to packet loss. The r reaction decreases the packet loss for ow r. To this point, we don t have any intuition of how a ow may decrease its packet loss when this exceeds certain bounds. To understand this, we need to recall that there are two main causes of packet loss: (a) wireless link errors; and (b) interference due to ow contention. In this study we only account for interference losses. In practice, in a friendly (as opposed to hostile) environment we expect 8 represent the uid ow dynamics of the threshold as follows: thrr = axc (t)P { c > r } bxc (t)P { c r } A. Results We now repeat the same simulation shown in Fig. 1 adding FairCast. We can see in Fig. 7 that the middle ow, ow two, increases its goodput from as little as 40% to 70%, thus achieving a 75% improvement. In this simulation the symmetry of the topology makes ow react at the sources. In general, we observe that in the case of xed routing FairCast is able to improve fairness between ows, redistributing channel access and leading spatially disgraced ows to achieve acceptable performance. The next step is to integrate FairCast with a routing protocol. This must be carefully done, since this interaction may lead to unexpected results. We here test FairCast over ODMRP. Recall that ODMRP builds a mesh routing path between a source and its receivers. When FairCast runs over ODMRP, different branches will achieve different delivery ratios, as expected in an adaptive layer scheme. In fact, this is an important feature of FairCast, namely, the ability to adjust individual branch rates to local bandwidth availability. This feature is essential in heterogeneous receiver scenarios. We also note that in Fig. 3,the overall delivery ratio for the video multicast application has decreased, while the audio rate has increased. This is a well known property of fairness - the total aggregate throughput is generally lower than that of the unfair solution. VI. CONCLUSION In a network, ows generally adapt to network feedback such as loss and delay. This paper shows that ows can actually interact and collaborate adapting to in-network signals and achieving a local fair share of resources. Past research on fairness in ad hoc networks only focused on unicast TCP ows and a number of optimizations have been introduced along the years to improve fairness and utilization of such protocol in wireless scenarios. We here present an algorithm by which it is possible to control the behavior, in a totally decentralized manner, between real-time multicast ows with adaptive layer adjustment/drop. FairCast ows are able to dynamically interact and acquire a fair share use of resources in a contention area. Future work will aim at: (a) providing FairCast s advantages for both elastic and inelastic traf c; (b) integrate FairCast with a number of routing protocols; (c) optimize the decentralized algorithm for threshold adaptation for multicast ows. (19) (20) At the steady state, averaging over time and assuming that thrr stabilizes in (0, 1) (i.e. boundaries excluded), we have that eq. (20) b (21) a+b Where c = E{ c (t)} and r = E{ r (t)}, the expected values of c (t) and r (t) at steady state. The values of a and b should re ect the fact that a higher rate of ow r generates higher values of c as a feedback from competing nodes. For this reason we use a heuristic that implements this concept, setting the values of a and b as follows: P { c > r } = a = k1 b = k2 (22) xout r (23) xin r where k1 and k2 are constants we set by testing a wide number of scenarios in simulation, xin is the incoming r rate per ow r and xout is the outgoing rate per ow r. r With the values set in eqs. (22) and (23) we have that at steady state P { c > r } = k2 xout /(k1 xin + k2 xout ). A r r r high relative value of xr induces a high probability that the received c is higher than r . In Fig. 7 we have an example of the performance of three FairCast ows, in the same scenario discussed in Fig. 1, when implementing the discussed dynamic threshold algorithm. As we can observe, both fairness and aggregate utilization are well preserved. V. EVALUATION We here show some preliminary results obtained using FairCast. In fact, it is possible to increase the number of ows that coexist by slightly lowering aggregate throughput and increasing fairness. The results are encouraging, indeed with a simple ow interaction protocol it is possible to improve the total QoS of the network. The results we here show have been obtained using a threshold adaptation algorithm in the unicast scenarios. Intuitively, a ow with a lower loss rate at a node has a higher average threshold, while a ow with a higher loss rate a lower average threshold. We instead use xed preset values in the multicast scenario. As we shall discuss in the following subsection, more work is needed to integrate FairCast and ODMRP. 9 Delivery Ration at 50 Pkt/s 1.0 1.0 FairCast Average Delivery Ration Stream 0 Stream 1 Stream 2 Receiver 1 Receiver 2 Receiver 3 Receiver 4 Receiver 5 Receiver 6 Receiver 7 Receiver 8 Receiver 9 Receiver 10 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 FairCast Fixed Routing FairCast 200Kbps FairCast 40Kbps Fig. 7. Unfairness between three ows, FairCast performance Fig. 8. Multicast scenario: two multicast ows, routed by ODMRP compared with other protocols and running FairCast, compete in the same area. Multicast ow one adapts, by lowering its rate, and lets multicast ow two in. REFERENCES [1] M. Allman, V. Paxson, and W. Stevens, TCP Congestion Control, in RFC 2581, April 1999. [2] M. Gerla, K. Tang, R. Bagrodia, TCP performance in wireless multi-hop networks, in Proceedings of 2nd IEEE Workshop on Mobile Computing Systems and Applications (WMCSA 99), February 25-26, 1999, New Orleans, Louisiana, USA. [3] C.E. Perkins, E.M. Royer, Ad-Hoc On Demand Distance Vector Routing, in Proceedings of the 2nd IEEE Workshop on Mobile Computing Systems and Applications (WMCSA), pp. 90-100, February, 1999, New Orleans, Alabama, USA. [4] D.B. Johnson, D. Maltz, Dynamic source routing in ad hoc wireless networks, in T. Imelinsky and H. Korth, editors, Mobile Computing, pp. 153-181. Kluwer Academic Publishers, 1996. [5] K. Xu, M. Gerla, L. Qi, Y. Shu, Enhancing TCP fairness in ad hoc wireless networks using neighborhood RED, in Proceedings of the Tenth Annual ACM/IEEE International Conference on Mobile Computing and Networking (MobiCom 03), pp. 1628, September 14-19, 2003, S. Diego, California, USA. [6] S. ElRakabawy, A. Klemm, C. Lindemann, TCP with Adaptive Pacing for Multihop Wireless Networks, in Proc. 6th ACM International Symposium on Mobile Ad Hoc Networking and Computing (MobiHoc 05) , Urbana-Champaign, IL, USA, May 2005. [7] S. Floyd, V. Jacobson, Random Early Detection gateways for Congestion Avoidance, in IEEE/ACM Transactions on Networking, V.1 N.4, August 1993, pp. 397-413. [8] H. Luo, S. Lu, V. Bharghavan, A New Model for Packet Scheduling in Multihop Wireless Networks, in Proceedings of the Sixth Annual ACM/IEEE International Conference on Mobile Computing and Networking (MobiCom 00), pp. 76-86, August 6-11, 2000, Boston, MA, USA. [9] J. Jun, M.L. Sichitiu, Fairness and QoS in Multihop Wireless Networks, in Proc. of the IEEE Vehicular Technology Conference (VTC 03), Orlando, FL, Oct. 6-9, 2003. [10] I. Gruber, A. Baessler, H. Li, Fair WLAN Scheduling for Ad Hoc Networks with Access Points, Poster, in ACM International Symposium on Mobile Ad Hoc Networking and Computing (MobiHoc 04), May 24-26, 2004, Tokyo, Japan. [11] A. Woo, D. Culler, A Transmission Control Scheme for Media Access in Sensor Networks, in Proceedings of the ACM/IEEE International Conference on Mobile Computing and Networking (MobiCom 01), pp. 221235, July 2001, Rome, Italy. [12] V. Bharghavan, A. Demers, S. Shenker, and L. Zhang, MACAW: A Media Access Protocol for Wireless LANs, in SIGCOMM Symposium on Communications Architectures and Protocols, pp. 212-225, September 1994, London, UK. [13] S.J. Lee, M. Gerla, and C. -C. Chiang, On Demand Multicast Routing Protocol, in Proceedings of IEEE Wireless Communications and Networking Conference (WCNC 99), pp. 12981302, September 1999. [14] S.J. Lee, W. Su, J. Hsu, M. Gerla, R. Bagrodia, A Performance Comparison Study of Ad Hoc Wireless Multicast Protocols , in Proceedings of the Nineteenth Annual Joint Conference of the IEEE Computer and Communications Societies of the IEEE (INFOCOM 00), March 26-30, 2000, Tel Aviv, Israel. [15] D. Palomar, M. Chiang, Alternative decompositions for distributed maximization of network utility: framework and applications, in Proc. IEEE Infocom, Barcelona, Spain, April 2006. [16] D. Bertsekas, Nonlinear Programming, Belmont, MA, Athena, 1995.
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:
inelastic_cac.pdf
Path: UCLA >> CS >> 465 Fall, 2008
Path: UCLA >> CS >> 465 Fall, 2008
Path: UCLA >> CS >> 467 Fall, 2008
Path: UCLA >> CS >> 461 Fall, 2008
Path: UCLA >> CS >> 461 Fall, 2008
Path: UCLA >> CS >> 460 Fall, 2008
Path: UCLA >> CS >> 463 Fall, 2008
Path: UCLA >> CS >> 463 Fall, 2008
Path: UCLA >> CS >> 399 Fall, 2008
Path: UCLA >> CS >> 393 Fall, 2008
Path: UCLA >> CS >> 392 Fall, 2008
Path: UCLA >> CS >> 391 Fall, 2008
Path: UCLA >> CS >> 395 Fall, 2008
Path: UCLA >> PSYCH >> 2305 Fall, 2008
Path: UCLA >> PSYCH >> 2305 Fall, 2008
Path: UCLA >> CS >> 147 Fall, 2008
Path: UCLA >> CS >> 317 Fall, 2008
Path: UCLA >> CS >> 120 Fall, 2008
Path: UCLA >> CS >> 122 Fall, 2008
Path: UCLA >> CS >> 126 Fall, 2008
Path: UCLA >> ATMOS >> 121 Fall, 1944
Path: UCLA >> ATMOS >> 121 Fall, 1944
Path: UCLA >> CS >> 204 Fall, 2008
Path: UCLA >> CS >> 206 Fall, 2008
Path: UCLA >> ATMOS >> 115 Fall, 1940
Path: UCLA >> ATMOS >> 115 Fall, 1940
Path: UCLA >> CS >> 453 Fall, 2008
Path: UCLA >> CS >> 447 Fall, 2008
Path: UCLA >> CS >> 446 Fall, 2008
Path: UCLA >> CS >> 445 Fall, 2008
Path: UCLA >> CS >> 444 Fall, 2008
Path: UCLA >> CS >> 443 Fall, 2008
Path: UCLA >> CS >> 442 Fall, 2008
Path: UCLA >> CS >> 441 Fall, 2008
Path: UCLA >> CS >> 440 Fall, 2008
Path: UCLA >> CS >> 449 Fall, 2008
Path: UCLA >> CS >> 448 Fall, 2008
Path: UCLA >> CS >> 372 Fall, 2008
Path: UCLA >> ATMOS >> 219 Fall, 1982
Path: UCLA >> ATMOS >> 219 Fall, 1982
Path: UCLA >> ATMOS >> 219 Fall, 1982
Path: UCLA >> ATMOS >> 219 Fall, 1982
Path: UCLA >> CS >> 270 Fall, 2008
Path: UCLA >> CS >> 271 Fall, 2008
Path: UCLA >> CS >> 148 Fall, 2008
Path: UCLA >> CS >> 142 Fall, 2008
Path: UCLA >> CS >> 141 Fall, 2008
Path: UCLA >> CS >> 144 Fall, 2008
Path: UCLA >> CS >> 145 Fall, 2008
Path: UCLA >> CS >> 145 Fall, 2008
Path: UCLA >> CS >> 482 Fall, 2008
Path: UCLA >> CS >> 487 Fall, 2008
Path: UCLA >> CS >> 479 Fall, 2008
Path: UCLA >> CS >> 479 Fall, 2008
Path: UCLA >> CS >> 476 Fall, 2008
Path: UCLA >> CS >> 477 Fall, 2008
Path: UCLA >> CS >> 474 Fall, 2008
Path: UCLA >> CS >> 472 Fall, 2008
Path: UCLA >> CS >> 470 Fall, 2008
Path: UCLA >> CS >> 470 Fall, 2008
Path: UCLA >> CS >> 471 Fall, 2008
Path: UCLA >> CS >> 389 Fall, 2008
Path: UCLA >> CS >> 384 Fall, 2008
Path: UCLA >> CS >> 383 Fall, 2008
Path: UCLA >> CS >> 193 Fall, 2008
Path: UCLA >> CS >> 192 Fall, 2008
Path: UCLA >> CS >> 227 Fall, 2008
Path: UCLA >> CS >> 226 Fall, 2008
Path: UCLA >> CS >> 223 Fall, 2008
Path: UCLA >> CS >> 221 Fall, 2008