MIMO AND TCP: A CASE FOR CROSS LAYER DESIGN

Transcription

1 MIMO AND TCP: A CASE FOR CROSS LAYER DESIGN Soon Y. Oh, Mario Gerla Computer Science Department, University of California, Los Angeles Los Angeles, CA USA {soonoh, gerla}@cs.ucla.edu and Joon-Sang Park Department of Computer Engineering, Hongik University Seoul, Korea jsp@cs.hongik.ac.kr Abstract TCP has not fared well in wireless ad hoc networks. Even in a perfectly static network, there is the problem of throughput degradation due to the interference between flows, which becomes more severe as the nodes come close to each other. Moreover, when two or more TCP flows compete in the same collision domain, often one of the flows captures the channel and blocks other flows, giving flows an unfair access to the channel. These problems have been traced in the MAC protocol and the interaction between TCP and MAC protocol retransmission timeouts. Proposed solutions range from modifications in the MAC protocols to the use of network layer selective drop strategies. However, none of these schemes offers full protection. Moreover, they require change in the existing protocols. The advent of Multiple-Input Multiple-Output (MIMO) system offers a way of increasing total throughput by reducing the interference. The same interference blocking feature leads to unfairness reduction. Recently proposed MIMO MAC protocols such as SPACE-MAC reduce interference by separating flows. In this paper, we test the efficacy of SPACE-MAC in eliminating the TCP capture problem and reducing throughput degradation due to interference. We evaluate the fairness and throughput performance of SPACE-MAC in various scenarios, and then compare with conventional MAC results via Qualnet simulation. I. INTRODUCTION Transport Control Protocol (TCP) is the most frequently used data transport protocol in Internet that allows reliable and congestion controlled end-to-end transmission. TCP continues to be the dominant transport protocol for the wireless ad hoc network technology based on the IEEE MAC protocol that has emerged. TCP, however, is reported to demonstrate poor performance in wireless ad hoc networks in [1], [2], [6], and [11] in terms of fairness and network-wise throughput. TCP s unfairness is triggered by many specific factors such as channel capture, hidden and exposed terminal conditions, and the binary exponential backoff mechanism of IEEE MAC, etc [1]. Throughput of the flows is also challenged in the presence of inter-flow interference. While these factors may be traced in the IEEE MAC protocol itself, TCP s greedy behavior and poor interaction with the MAC layer further boost the unfairness and throughput degradation. Proposed solutions range from modifications in the MAC protocols to the use of network layer selective drop strategies. The objective is to equalize the aggressiveness of the flows while maintaining the throughput, yet none of the solutions succeeds at both ends. Besides, they are not very attractive as they require radical changes in protocols. The advent of Multiple-Input Multiple-Output (MIMO) system [3] offers an attractive way of resolving the conflicts between the flows through interference reduction. We can carefully tune the MIMO antenna weights of the transmitter and receiver pairs to limit the interference between TCP flows by separating them, thus improving fairness and throughput. Recently proposed MIMO MAC protocols [5], [7], [8] perform such flow separation and are proven to be very effective in various simulation experiments. In this paper, we test the efficacy of MIMO MAC space separation in eliminating the TCP capture problem and in maintaining reasonable performance. In this paper, we evaluate the fairness of SPACE-MAC [7], a MAC protocol that leverages the advantages of MIMO system, under varying TCP offered loads, in different configurations and in various scenarios via Qualnet simulations. The SPACE-MAC results are compared with conventional IEEE MAC results. The rest of paper is organized as follows. We review the problems of TCP unfairness and performance in section 2 and describe in short the SPACE-MAC protocol in 1 of 6

2 section 3. Section 4 shows simulation environments and results. We conclude the paper in section 5. II. TCP FAIRNESS AND PERFORMANCE IN WIRELESS AD HOC NETWORKS Spatial reuse and range dependency are the factors that influence TCP s fairness and performance in wireless ad hoc networks. The former is due to spectrum limitation. If two or more TCP flows are present in the same collision domain, they compete for shared spectrum and interfere with each other. Often one of the flows wins the competition and captures the channel at the expense of other flows. The latter, range dependency, is often recognized as the primary reason for TCP unfairness. It refers to the fact that a receiver can receive and decode the signal transmitted by a transmitter within the range R 1, and cannot decode but is still interfered by the signal transmitted by a transmitter within the range R 2, where R 1 < R 2. We call R 1 the reception range and R 2 the interference range. In Qualnet, for example, the reception range and the interference range are set to 37m and 7m, respectively, by default. This range dependency causes various problems including hidden/exposed terminal problems. Packets may collide and be dropped due to interference that cannot be detected a priori. When a node fails to reach its neighbor due to interference from other nodes, MAC triggers binary exponential backoff. After failing to reach a neighbor node several times, the transmitter node considers this as a route failure and starts route discovery process. If this persists, TCP will also reach timeout and invoke its own congestion control algorithm, further delaying the retransmission of packets. Flow 3 Node C Node B Flow 2 Node A Flow 1 Fig 1. Different channel access opportunity due to topology A topology may play a role as well. Nodes in a wireless ad hoc network have unequal channel access due to the topology. A node that competes with fewer flows gets higher opportunity to access the channel and better throughput than a node with more interfering flows in its range. Fig 1 illustrates an example. Suppose there are three TCP flows, and the distances between Nodes A and B and Nodes B and C are both 4m. This means the sources and destinations of the flows are placed outside other nodes reception ranges, but the intermediate Node B is within the interference ranges of Nodes A and C. Therefore, Flows 1 and 3 have only one interfering flow each, while Flow 2 has two interfering flows. Intuitively, Flow 2 achieves lower throughput because of higher contention. We simulated the above scenario in Qualnet, and Fig 2 graphs the throughput of three flows, performed in two different environments. We first run the simulation assuming wireless channel with no fading, and run again with Rayleigh fading, to see the effect of channel model. For both cases, we use TCP RENO with send and receive buffers with the sizes of 32 packets each. The reception range is 37m and the distance between Nodes A and B and Nodes B and C are 4m. Under no fading, Flow 2 is completely starved while Flow 1 and 3 achieve over 4Kbps of throughput. Flow 2 packets repeatedly collide and are dropped at the Node B, driving the transmitter of Flow 2 to increase timeouts and backoff intervals. This result coincides with the Xu s previous simulation [1], and demonstrates the extreme case in which the flows get unfair access to the channel. With the channel that exhibits Rayleigh fading, though, Flow 2 manages to deliver some packets, mainly due to the randomness that Rayleigh fading introduces. Still, the fairness problem does not go away, as we see Flow 2 suffering from the competition. Even worse, we observe significant throughput degradation, in part introduced by Rayleigh fading itself and by TCP reacting against Flow 2 s intervention. The throughput difference suggests that the choice of channel model is yet another factor that determines the outcome of our study. No Fading vs. Rayleigh Fading Throughput (Kbits/s) Flow 1 Flow 2 Flow No Fading Fig 2. Overall throughput of flows in Fig 1 Rayleigh Fading 2 of 6

3 III. SPACE-MAC SPACE-MAC is a MAC protocol that leverages the advantages of MIMO system. Before describing the SPACE-MAC protocol, we present the basic MIMO model. We assume that the channel exhibits flat fading and channel side information is known both at the sender and the receiver. We also assume rich scattering environment. For the following discussion, we use uppercase and lowercase boldface letters to denote matrices and vectors, respectively, and superscripts T and H to denote the transpose and Hermitian operations, respectively. MIMO technology features two key mechanisms, the first of which is known as spatial multiplexing [4]. Under spatial multiplexing, MIMO channel can be decomposed into as many as N parallel independent channels. Therefore, the input signal from the upper layer is split into N independent signals and delivered via these multiplexed channels. This results in an N-fold increase in data rate comparing to single antenna systems, where N is bounded by the lesser of number of antennas at the transmitter and receiver. The second feature is MIMO beamforming [4], which achieves diversity gain by maximizing the signal strength, instead of trying to send many data streams concurrently. This involves appropriately weighting signals both at the transmitter and at the receiver. At the transmitter, the same signal is weighted independently before being fed into different antennas, and at the receiver, coherent combining of the beams is performed, such that the signal-to-noise ratio (SNR) is optimized, without boosting transmit signal power or bandwidth. SPACE-MAC performs the combination of two techniques and exploits MIMO and achieves spatial reuse by allowing the concurrent exchange of data by multiple transmitter/receiver pairs in the same collision domain, as shown in Fig 3. SPACE-MAC allows Node A to transmit a packet to Node C while Node B is transmitting a packet to Node D. Each node protects the transmission and reception of its signal by selectively nulling the interfering signals. This is enabled in a completely distributed way by adjusting antenna weights, given channel state information. In SPACE-MAC, a transmitter sends an RTS using its weight vector, w T. The same weight vector, w T will be used to send the next data packet and to receive the corresponding CTS and ACK. When the designated receiver receives the RTS, it responds with a CTS packet using its weight vector, w R, which will be used to receive the data packet and to send ACK. The nodes that overhear this RTS or CTS estimate the effective channel and adjust their weight vectors such that they can nullify the signal from the transmitter. After this RTS/CTS exchange, the transmitter sends and the receiver receives a data frame using their weight vectors. The nodes that are able to nullify these signals can transmit and receive their own signals without interfering with the existing ones. Node A Node B Node S Node D Node C Fig 3. Simple topology in which SPACE-MAC achieves spatial reuse Consider the topology illustrated in Fig 3. Suppose Node A chooses its weight vector w A and starts transmission. The designated receiver, Node C, estimates the channel vector h AC = w A H H AC, where H AC is the channel matrix between A and C with each element being the antennato-antenna channel gain, and sets its weight vector to w C = h AC T, which maximizes the combined channel and diversity gain. Now suppose Node B wants to transmit a packet to Node D. B performs the nullifying procedure by adjusting its weight vector w B, such that it satisfies w A H H AB w B = and w C H H CB w B =, so that B does not disturb the transmission between A and C. Node D should have overheard A s RTS and C s CTS, thus have adjusted its weight vector to nullify the signal from A to C. D can choose to use its current weight vector or some new w D that maximizes the effective channel gain from B while nullifying other signals. To select a new weight vector, D must have enough Degrees of Freedom (DOF), which is essentially the number of antennas it has. Nullifying of a single signal consumes one DOF, which means that an N-antenna system can null out at most N-1 nodes. Any additional new transmission is possible only when both transmitter and receiver have enough degrees of freedom. In the case depicted by Fig 3, for the 4 nodes A, B, C, and D to transmit simultaneously, each node must have at least 4 nodes, because they need to nullify 3 signals. If there is a Node S that shares the channel with these 4 nodes, though, it cannot start a new transmission because it has used all its 4 DOFs to nullify the 4 data streams. 3 of 6

4 IV. SIMULATION RESULTS In this section, we discuss our simulation setups and results. We compare the performance of SPACE-MAC to IEEE MAC in different scenarios. SPACE-MAC is implemented as explained in the previous section in Qualnet [9]. We add data structures that hold antenna weights and channel information. The antenna weights are calculated on the fly and are applied to the computation of the SNR. We assume that the channel estimation can be done with 82.11a/b/g PLCP preamble and that there is no error in the channel estimation. We use the default values that Qualnet provides for the configurable parameters unless otherwise specified. Each node is modeled to be equipped 5 antennas each. Our simulations assume a quasi-static Rayleigh fading environment. In other words, the channel is invariant during a complete communication session including RTS, CTS, DATA, and ACK frame exchanges. The channel coefficient for a transmit/receive antenna pair is modeled as a complex Gaussian random variable with zero mean and unit variance, statistically independent of the other coefficients. We use TCP RENO with maximum segment size set to 512 bytes and the send and receive buffers with the sizes of 32 packets each. Static routing is used and TCP connections start at T s = 1 second and end at T e = 13 second, while the simulation duration is 15s. The channel bandwidth is 2Mbps and reception range for all nodes is 37m. We use 5-antenna MIMO in our simulation, for hardware limitation usually will not allow more than 5 antennas to be equipped on a laptop or a mobile device. We run simulations of FTP flows with each TCP packet being 512 bytes long and average values obtained from ten simulation experiments for each scenario. We evaluate throughput and fairness, where the throughput is defined as the total received bytes of the designated receiver divided by total time of the transmitter sending the packets, which is 12 seconds in our simulation. the slow start period, a TCP sender increases its congestion window exponentially by doubling its size every Round Trip Time. Thus the TCP sending rate at the sender continues to increase exponentially. SPACE-MAC achieves spatial reuse via weight vector adjustment. If a node receives a CTS packet from another flow, it picks its weight vector that would not disturb existing communication and starts its transmission. When the size of congestion window increases, a SPACE-MAC transmitter tries to transmit packets sequentially. Even though the source hears a CTS packet from an intermediate node (that CTS is sent for a destination receiver), it transmits an RTS or a DATA packet after adjusting a weight vector. This continuous packet stream is collided at the intermediate node with a TCP ACK from a designated receiver since the intermediate node cannot receive both signals simultaneously. Hence, as the congestion window increases, SPACE-MAC throughput decreases. We ran simulation experiments with various congestion window sizes and found that in this particular scenario SPACE-MAC yields the best throughput when the congestion window is fixed to 1. Thus, in the following simulations, we fixed the maximum congestion window size to 1 in TCP RENO for both MAC and SPACE-MAC. Before we actually discuss our simulation results, let us present the result of Xu s experiment [1] in Fig 4, where he considered the very same scenario where no fading was assumed to evaluate Neighborhood RED (NRED) scheme. The set of graphs on the left shows the throughput of each flow without NRED, and the ones on the right graphs the throughput of each flow with NRED. We can see that NRED mitigates the unfairness problem, while giving up some portion of the aggregated throughput. A. 3 FTP/TCP flows Now we consider the scenario of 3 parallel FTP/TCP flows, illustrated in Fig 1. Transmitters and receivers are located out of other s transmission range and they communicate with only intermediate nodes Nodes A, B, and C. The distance between Nodes A and B and Nodes B and C are 4m, and our Qualnet model assumes the reception range of 37m and the interference range of 7m. Fig 4. Xu s experiment on NRED s effect on TCP fairness Let us discuss the optimal TCP window size in a multihop case. TCP congestion control is very aggressive. In 4 of 6

5 Throughput (Kbits/s) FTP/TCP Flows Flow 1 Flow 2 Flow SPACE-MAC MAC Protocol Fig 5. The throughput of 3 FTP/TCP flows with the distance between flows being 4m Now we present the result we have obtained from the simulation runs where we assume wireless channel that exhibits Rayleigh fading in Fig 5, where we show the throughput of each of three FTP/TCP flows. We observe that MIMO with SPACE-MAC helps reduce the unfairness issue that is present in MAC s case, while also mitigating the throughput degradation that Rayleigh fading has introduced (see Fig 2). Flow 2 manages to deliver some packets to the destination despite the interference from Flow 1 and Flow 3 in SPACE-MAC s case, because multiple antennas at each node inherently provide some degree of diversity gain. Furthermore, because the weights are initially selected in a random fashion before nodes make contact with neighboring nodes, there is some chance for Flow 2 s RTS to penetrate the interference and reaches Node B. More precisely, if weights are selected such that A and C s interference at B is weak and the sender of Flow 2 s signal is felt strong at B, there may be some packets that survive the interference, and the randomness raises the chance of this happening. Once an RTS reaches B that way, B can estimate the channel and correspondingly adjust its weights to maximize the received signal SNR, further increasing the chance of the successful packet transfers. This accounts for the improved throughput for Flow 2 in SPACE-MAC s case in the simulation, and therefore achieves better fairness than s case. The juxtaposition of Xu s work and ours help us conclude that SPACE-MAC is equivalent of NRED scheme under the presence of Rayleigh fading. A careful reader may have noticed that in Fig 1 scenario we re-establish fairness by simply improving the received signal strength at Node B via MIMO (thus in part overcoming the interference). However, B cannot utilize the nulling feature of SPACE-MAC since Nodes A and C are out of B s reception range, i.e. their RTS and CTS packets cannot be heard by B. To allow the nulling property to take an effect, we have slightly modified the Fig 1 topology reducing A and C distance to B to 35 meters. In this case, two things happen at the same time: B is in the reception ranges of A and C, while A and C interfere with each other. This is the extreme case where all three flows contend for the channel only one flow can proceed at a time. Throughput (Kbits/s) We show the results for distance D = 35m in Fig 6. Having all three flows contend for the channel, MAC recovers fairness, mostly by reducing the throughput of Flows 1 and 3. SPACE-MAC, on the other hand, allows Flow 2 to null the unwanted signals from A and C, which are within its reception range, and retains high throughput while solving the fairness problem. The total throughputs are compared in Fig 7. This big throughput advantage is due to 1) the nullifying feature of SPACE-MAC, that allows the three flows to be supported in parallel, 2) the inherent diversity gain that the array of antennas introduces, and 3) random initial assignment of antenna weights and SNR optimization, which allows packets to penetrate the interference and noise, as in the 4m case. B. More Realistic Scenario 3 FTP Flows Flow 1 Flow 2 Flow Space-MAC MAC Protocol Fig 6. The throughput of 3 FTP/TCP flows with the distance between flows being 35m Throughput (Kbit/s) Total Throughput Space-MAC Fig 7. The aggregated throughput of the flows with D=35m In the previous experiment, the topology was simple and was specifically chosen to illustrate the problems with the 5 of 6

6 82.11 MAC and the benefits of the SPACE-MAC. In the next set of experiments, random topology and traffic are chosen to implement more realistic scenarios. 1 nodes are uniformly distributed in 75m by 75m area and 2 FTP/TCP connections are randomly selected. Simulation time is 15 seconds and 2 transmitters start transmission at the same time, T s =1 second, and end at T e =13 second. From Fig 8 and Fig 9, we observe that SPACE-MAC performs better in terms of total throughput. The severe interference among the flows hurts all flows, similar to the 35m case we presented above. In SPACE-MAC better throughput has been achieved compare to MAC because the interference is mitigated at each node. This, while not as much as in the 35m case, gives some performance advantage to SPACE-MAC over The nodes still suffer from some interference because the number of flows is greater than the number of antennas that each node has, i.e. a node cannot nullify all the interfering flows. Throughput (Kbits/s) Random 2 Flows Space-MAC Recievers Fig 8. The individual throughput of 2 random flows Throughput (Kbit/s) Total Throughput SPACE-MAC Fig 9. The total throughput of the flows V. CONCLUSION TCP is known to perform poorly in wireless ad hoc networks that use IEEE MAC protocol both in terms of throughput and fairness. In this paper, we examined MIMO technology as the solution to the problem, using SPACE-MAC as the MAC protocol. A series of experiments indicates that MIMO, when used wisely, can help reduce the effect of interference at each node to help improve the TCP performance. This reveals that there is a high correlation between the operations of TCP and lower layered MIMO architecture, suggesting a cross-layer solution to improve the performance of ad hoc networks. ACKNOWLEDGEMENTS 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-6-3-1, and; by ARMY MURI under funding W911NF We would like to thank Brain Choi, a PhD student of University of California, Los Angeles for his valuable review and revision of our paper. References [1] Z. Fu, H. Luo, P. Zerfos, L. Zhang, and M. Gerla. The Impact of Multihop Wireless Channel on TCP Performance. IEEE Transactions on Mobile Computing Vol.4 Mar. 5. [2] M. Gerla, K. Tang, and R. Bagrodia. TCP Performance in Wireless Multi-hop Networks. Proceedings of IEEE WMCSA 99, Feb [3] D. Gesbert, M. Shafi, D. Shiu and A. Naguib. From Theory to Practice: An Overview of MIMO Space-time Coded Wireless Systems. IEEE Journal of Selected Areas in Communications Vol 21, No 3, 3. [4] A. Goldsmith. Wireless Communications. Cambridge University Press, 5. [5] J. Mundarath, P. Ramanathan, and B. D. Van Veen. NULLHOC: A MAC Protocol for Adaptive Antenna Array Based Wireless Ad Hoc Networks in Multipath Environments. Proceedings of the IEEE Global Telecommunications Conference (Globecom), 4. [6] D. Perkins and H. Hughes. TCP Performance in Mobile Ad Hoc Networks. Proceedings of IEEE WCNC, [7] J.S. Park, A. Nandan, M. Gerla, and H.Lee. SPACE-MAC: Enabling Spatial Reuse using MIMO channel-aware MAC. Proceeding of the IEEE International Conference on Communications (ICC), 5. [8] M. Park, S-H Choi, and S. Nettles. Cross-layer MAC Design for Wireless Networks Using MIMO. Proceeding of the IEEE Global Telecommunications Conference (Globecom), 5 [9] Scalable Networks Inc. QualNet. [1] K. Xu, M. Gerla, L. Qi, and Y. Shu. Enhancing TCP Fairness in Ad Hoc Wireless Networks Using Neighborhood RED. Proceedings of ACM MobiCom 3, 3. [11] S. Xu, and T. Saadawi. Does the IEEE MAC Protocol Work Well in Multihop Wireless Ad Hoc Networks? IEEE Communications Magazine, 39(6), Jun of 6