Effects of Ad Hoc MAC Layer Medium Access Mechanisms under TCP

Transcription

1 Mobile Networks and Applications 6, , Kluwer Academic Publishers. Manufactured in The Netherlands. Effects of Ad Hoc MAC Layer Medium Access Mechanisms under TCP KEN TANG, MARIO CORREA and MARIO GERLA Computer Science Department, University of California, Los Angeles, USA Abstract. Mobile computing is the way of the future, as evident by such initiatives as Bluetooth, Iceberg and HomeRF. However, for mobile computing to be successful, an obvious layer, the MAC layer, must be efficient in channel access and reservation. Therefore, in-depth understanding is needed of the wireless MAC layer if wireless computing is to takeoff. Many random access wireless MAC protocols have been proposed and even standardized. However, there has yet been an attempt to understand why certain designs are used and what makes certain protocols better than others. In this paper, we survey several popular, contemporary, wireless, random access MAC protocols and determine the effects behind the design choices of these protocols. Keywords: wireless, ad hoc, random access, MAC 1. Introduction Mobile computing has proliferated in recent years. Initiatives, such as Bluetooth [6], Iceberg [8] and HomeRF [11], are making wireless networking a reality. Such technologies allow for wireless communications between numerous devices, such as the PC, PDA, cellular and cordless phones, pagers and other appliances, to be practical and seamless. With such diverse and dense communication environment, controlling access to the wireless medium plays an integral part in making Bluetooth, Iceberg and HomeRF become reality. Therefore, the success of mobile computing technologies will in part hinge on the effectiveness of the wireless MAC layer protocol being deployed. In this paper, we provide insights into the design of contemporary wireless, ad hoc random access MAC layer protocols. In particular, we focus on the various methods of channel access used to resolve contention on the wireless medium. We isolate individual medium access mechanisms (such as carrier sensing, RTS/CTS control frames and linklevel ACKs) through experiments to determine the effect each mechanism has on the network. In section 2, an overview of several random access wireless MAC layers is introduced. Section 3 discusses our research study. Simulation experiments are described in section 4. Results from our UDP and TCP experiments under various MAC protocols are presented in sections 5 and 6, respectively. Finally, section 7 concludes the paper. 2. Existing wireless random access MAC schemes This section presents an overview of several popular random access wireless MAC layer protocols. The protocols include Carrier Sense Multiple Access (CSMA), Multiple Access with Collision Avoidance (MACA), MACAW, Floor Acquisition Multiple Access (FAMA) and IEEE CSMA (Carrier Sense Multiple Access) requires carrier sensing before transmission. If the channel is free, the packet is transmitted immediately. Otherwise, it is rescheduled after a random timeout. The major limitation of CSMA is the hidden terminal and exposed terminal problem [12]. CSMA was used first in the Packet Radio network in the mid 1970 s [9]. An alternative to this scheme involves channel reservation overhead in order to reduce the number of collisions. This scheme, called Multiple Access with Collision Avoidance (MACA), utilizes control frames, which can be overheard by stations with the ability to collide with the upcoming transmission [10]. These control frames are of fixed size, and also contain the length of the proposed data transmission. The station wanting to transmit initiates a Request To Send (RTS) control frame, and sets a timer to await the Clear To Send (CTS) control frame from the intended receiver. Any station overhearing the RTS control frame is within range to collide with the expected CTS frame and must defer transmission for a period of time long enough to ensure delivery of the CTS. If the station does not receive the CTS within a specified timeout interval, it assumes a collision has occurred and initiates a backoff before attempting to retransmit. Any station overhearing the CTS control frame is within range to collide with the upcoming transmission and is required to defer for the full duration of transmission, as specified in the control frame. MACAW [2] is an enhancement to MACA. MACAW utilizes a MILD backoff to select the retransmission time. Besides using RTS/CTS, three other control frames are introduced (DS, ACK, RRTS). The protocol also implements a backoff copying scheme that allows stations to use the same backoff counters when transmitting to the same station. Moreover, a multiple stream model is included to foster fairness. Floor Acquisition Multiple Access (FAMA) is an experimental MAC protocol specifically developed for the Glomo

2 318 K. TANG ET AL. DARPA program [4]. It is a descendent of MACA. Like MACA, it uses RTS and CTS exchange to prepare the floor (medium) for data transmission. However, unlike MACA, it also utilizes carrier sensing. IEEE [7] expands on FAMA by adding link-level ACKs in addition to the RTS/CTS control frames. Furthermore, abandons the traditional CSMA in favor of CSMA/CA (Collision Avoidance). In , the Distributed Coordination Function (DCF) represents the basic access method that mobile nodes utilize to share the wireless channel. The scheme incorporates CSMA with Collision Avoidance (CSMA/CA) and acknowledgement (ACK). Optionally, the mobile nodes can make use of the virtual carrier sense mechanism that employs RTS/CTS exchange for channel reservation and fragmentation of packets in situations where the wireless channel experiences high bit error rate. CSMA/CA works as follows: a node wishing to transmit senses the channel, and if it is free for a time equal to the DCF InterFrame Space (DIFS), the node transmits. If the channel is busy, the node enters a state of collision avoidance and backs off from transmitting for a specified interval. In the collision avoidance state, the node sensing the channel busy will suspend its backoff timer, only resuming the backoff countdown when the channel is again sensed free for a DIFS period. A typical sequence of exchange in using the virtual carrier sensing mechanism involves the source node first sensing the channel using CSMA/CA. After CSMA/CA is executed, the source node transmits a RTS, followed by the destination node responding with a CTS, then with the source node sending the data frame and ending with the destination node confirming with an ACK to the source node. 3. Research goals The various random access schemes for wireless ad hoc networks presented in section 2 employ different methods, or mechanisms, for channel contention on the wireless medium. Among them are carrier sense, packet sense (no carrier sensing), carrier sense with collision avoidance, RTS/CTS control frames and link-level ACKs. For example, CSMA solely uses carrier sense. FAMA utilizes carrier sense with RTS/CTS control frames. MACA, on the other hand, digresses from the carrier sensing of FAMA and instead opts for packet sensing. MACAW adds on top of MACA link-level ACKs, among other features. Finally, (with virtual carrier sense) coalesces FAMA, linklevel ACKs and collision avoidance. In our study, we isolate each of the above features to determine the individual effect each mechanism incurs on network performance and at the same time determine which combination works best under UDP and TCP. In particular, we hope to answer the following questions: First of all, what kind of basic medium access technique works best? Is it PSMA, CSMA or CSMA/CA? Also, it is well known that RTS/CTS improves the system performance of wireless networks. However, does the improvements aid in fairness, aggregate throughput or both? Furthermore, link-level ACKs have been suggested to alleviate the problems associated with the high error rates of the wireless channel. To what extent are link-level ACKs useful? Finally, with many possible combinations of wireless access mechanisms, which combination works best? Although there are many important metrics to consider, in this paper we focus on throughput and fairness. For our study, we implemented various combinations of the above features and compare them to one another in an effort to isolate each mechanism. The protocols implemented are packet sense (PSMA), carrier sense (CSMA), carrier sense with collision avoidance (CSMA/ CA), PSMA/RTS/CTS, CSMA/RTS/CTS, CSMA/CA/RTS/ CTS, PSMA/ACK, CSMA/ACK, CSMA/CA/ACK, PSMA/ RTS/CTS/ACK, CSMA/RTS/CTS/ACK and CSMA/CA/ RTS/CTS/ACK. The simulation platform used is GloMoSim [13]. Glo- MoSim is a discrete event parallel simulation environment implemented in PARSEC, PARallel Simulation Environment for Complex Systems [1]. It includes various detailed, wireless protocols in its library (radio propagation, mobility, MAC, network, transport and applications). In addition, GloMoSim provides a valuable and useful feature that facilitates different protocols at a given layer to be swapped in and out of the protocol stack and thus allows for comparison between these different protocols. Most importantly, Glo- MoSim permits the detailed modeling of several layers and the study of their interaction, yet preserving very good runtime efficiency and yielding manageable execution time. 4. Experimental configuration and parameters For our simulation experiments, we consider several topologies (figures 1 5): string, hidden terminal, exposed terminal, ring and bottleneck. The arrows represent the direction of data packet transmissions. Constant bit rate (CBR) with unbounded backlog executing with UDP or FTP with infinite backlog running on top of TCP is used for the application. We utilize static routing to route packets since mobility is not considered. Different MAC protocols are evaluated: PSMA, CSMA, CSMA/CA, PSMA/RTS/CTS, CSMA/RTS/CTS, CSMA/CA/RTS/CTS, PSMA/ACK, CSMA/ACK, CSMA/ CA/ACK, PSMA/RTS/CTS/ACK, CSMA/RTS/CTS/ACK and CSMA/CA/RTS/CTS/ACK. Each MAC protocol uses exponential backoff. Radios with no capture ability are modeled with a channel bandwidth of 2 Mbps. Furthermore, the channel uses free-space with no external noise (perfect channel). No capture ability of the radios and free-space channel model are used for easier understanding of the results due to their deterministic behavior as opposed to radios with capture ability and fading channel model, where there are randomness involved. Transmission and propagation delays are modeled. Processing delay is negligible. Each node has a 25-packet MAC layer buffer pool. Scheduling of packet transmissions is FIFO. The TCP simulation code was ported directly from FreeBSD code

3 EFFECTS OF AD HOC MAC LAYER MEDIUM ACCESS MECHANISMS 319 (TCP-Lite) into GloMoSim. TCP packet length is assumed fixed at 1460 B. In our experiments, we force the maximum TCP advertised window to be one packet size (i.e B) in order to better understand and analyze the network behavior. Also, it is shown in [5] that having an adaptive window does not help but rather worsens TCP s performance under a wireless ad hoc environment. TCP connections are started uniformly, distributed between 0 to 10 s. Each simulation run is executed for 200 simulated seconds. Five different sets of experiments are examined. First, with the variable number of hops experiment, we vary the number of hops for a connection from 1 to 5 using the topology in figure 1. Each node is only within range of their intermediate neighbors. This is an experiment useful to determine the effectiveness of each protocol when no contention exists. Next, we divert our attention to the hidden terminal environment depicted in figure 2. Node 1 is in radio range of node 0 and node 2. Node 0 and node 2 are not within reception range of each other. Connections are set up from node 0 to node 1 and from node 2 to node 1. With the ex- Figure 1. String topology. posed terminal experiments shown in figure 3, node 0 is in range of node 1, node 3 is in range of node 2, node 1 is in range of both node 0 and node 2 and node 2 is in range of both node 1 and node 3. Two connections are established, one between node 1 and node 0 and the other between node 2 and node 3. Then, we consider a ring topology as depicted in figure 4 to determine each protocol s fairness characteristics. The 8 nodes are engaged in single hop connections (node 0 to node 1, node 1 to node 2,..., node 7 to node 0) with each node being able to reach its neighboring nodes only. Finally, we examine multiple flows sharing a common bottleneck, as shown in figure 5. Here, we have connections from node 0 to node 5, node 1 to node 6, node 2 to node 7, node 3 to node 8 and node 4 to node 9. All sessions travel through common nodes and links, with the number of hops for the bottleneck link varying from 0 to 2. Furthermore, the senders (node 0, 1, 2, 3 and 4) are within transmission range of each other. The receivers (node 5, 6, 7, 8 and 9) are also within range of each other. The bottleneck experiments are used to gauge the behavior of the various protocols when contention exists for a particular link. In an attempt to shorten the length of the paper, we will use only tables to capture the abundant results from our simulation experiments. 5. Constant bit rate with UDP Figure 2. Hidden terminal topology. Figure 3. Exposed terminal topology. Figure 4. Ring topology. To investigate the impact of the individual medium access mechanisms (PSMA, CSMA, CSMA/CA, RTS/CTS and ACK) under UDP, we evaluate various protocols using the experiments mentioned in section 4 in addition to others. Here, the traffic constant bit rate is such that there will always be some packets to send from the queue. In other words, each node will have infinite backlog. The results of our experiments are presented in [3]. In this section, we provide a brief overview of the conclusions we derived in [3]. In [3], the results show that link-level acknowledgements do not contribute to the fairness of the various protocols. Furthermore, CSMA/RTS/CTS is the only protocol where ACKs consistently improves performance since they improve the coordination between sender and receiver. Overall, ACKs increase overhead and increase the possibility of collision. Thus, link-level ACKs is counterproductive under UDP since it contradicts the intent of UDP, which is to Figure 5. Bottleneck topology.

4 320 K. TANG ET AL. support unreliable transfer of data. RTS/CTS is responsible for the fairness (i.e. node coordination) of the protocols that employ it. By being fair and allowing all clients to successfully transmit packets, the RTS/CTS mechanism increases the throughput of each client/server pair. In summary, the degree of coordination is greater when using RTS/CTS than when using ACKs. When communicating using unreliable transmission it is desirable to be as aggressive as possible and have some form of coordination between nodes. Of the three basic protocols (PSMA, CSMA, and CSMA/CA), CSMA is the best building block over the topologies evaluated. It performs the best in the variable hops experiments. In the string experiments, CSMA/RTS/CTS/ACK is the only protocol that allows both clients to successfully send packets. CSMA/RTS/CTS performs better than all the other protocols and has a slight advantage over CSMA/CA/RTS/CTS and CSMA/CA/RTS/ CTS/ACK. Any protocol using RTS/CTS control frames, except PSMA/RTS/CTS/ACK, does well in the hidden terminal experiments and none of them has a clear advantage. The performance of CSMA is comparable to that of PSMA in the exposed terminal topology. In the more complex cross and bottleneck topologies, CSMA/RTS/CTS/ACK performs well. The protocol performs better over both cross traffic scenarios and its performance is comparable to that of PSMA/RTS/CTS in the bottleneck experiments. While the performance of CSMA/CA/RTS/CTS/ACK is comparable to that of CSMA/RTS/CTS/ACK in many topologies, it is a bit too conservative for this environment. The extra waiting time before transmission degrades performance in many of the more congested environments. 6. File transfer with TCP To determine the affects of the individual medium access mechanisms (PSMA, CSMA, CSMA/CA, RTS/CTS and ACK) under TCP, we contrast various protocols using the experiments mentioned in section 4. The traffic type is FTP with infinite backlog at each source node PSMA, CSMA and CSMA/CA We start by examining the performance of TCP on top of PSMA, CSMA and CSMA/CA with the variable number of hops experiments. Here no contention exists, and therefore, intuitively, we expect PSMA and CSMA to exhibit similar network behavior since carrier sensing is not needed. CSMA/CA, meanwhile, produces less throughput due to the unnecessary collision avoidance overhead. Table 1 confirms our insight. The throughput behavior shown in table 1 also supports the send-and-wait model (since TCP window size is set to one packet size). Assuming that propagation, processing and carrier-sensing delays are negligible, the expected utilization of a onehop connection for a send-and-wait protocol is given by td tx + ta, Table 1 Variable number of hops experiment, throughput in bits per second. Number of hops PSMA CSMA CSMA/CA Table 2 Hidden terminal experiment, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA where td, tx and ta are the transmission delay of actual application level data unit (1460 B), application level data unit with lower level headers and TCP ACK with lower level headers, respectively. With the size of TCP/IP header being 40 B and CSMA header being 12 B, the expected utilization of a one-hop connection is Thus, the expected throughput for a 2 Mbps channel is 1868 Kbps. For a twohop connection, we have td 2tx + 2ta. Here, the expected link utilization is 0.467, and thus, the expected throughput is 934 Kbps. The analytical results are slightly higher than simulation since propagation, processing, and carrier-sensing delays are not taken into consideration in the analytical model. The equation can be easily generalized to the n hop case. We note a very close match between simulation and analysis in table 1. Now, let us refer to the hidden terminal experiments. All three protocols (PSMA, CSMA and CSMA/CA) mirror one another (table 2). The connection from node 2 to node 1 captures the channel and locks out the connection from node 0 to node 1 completely. The capture behavior is explained by the interplay of MAC timeouts with TCP timeouts, coupled with the presence of undetected link losses. When two sessions compete for the same channel and one is pushed back by the timeouts, the binary backoff nature of both MAC and TCP timeouts makes the situation progressively worse for the loser. Note that the collision avoidance scheme of CSMA/CA alone does not work well without feedback from control frame losses, such as RTS or ACK frames. This is because the MAC backoff value does not increase when collisions of data frames due to contention occur since the MAC layer has no way of finding out about the collisions. Therefore, the collision avoidance part of CSMA/CA has no beneficial effect. In terms of throughput, CSMA outperforms PSMA and CSMA/CA. In the exposed terminal experiments (table 3), the connection between node 1 and node 0 is totally locked out in PSMA. This is counterintuitive since we expect a MAC layer protocol that does not carrier sense to do well in the exposed

5 EFFECTS OF AD HOC MAC LAYER MEDIUM ACCESS MECHANISMS 321 Table 3 Exposed terminal experiment, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA Table 4 Ring experiment, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA Table 5 Bottleneck experiment, zero hops, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA Table 6 Bottleneck experiments, one hop, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA terminal scenario [10]. However, since TCP acknowledges every data packet, the TCP ACK may deter node 0 from sending TCP ACK to node 1 since node 1 cannot receive any data packets from node 0 while node 2 is transmitting to node 3. What makes things worse is the fact that since the transmission delay of TCP ACKs are minimal compared to the data packets, the data packets from node 2 will more likely interfere with the ACKs sent by node 0 node 1. This scenario causes TCP ACKs from node 0 to be lost, resulting in retransmissions by node 1. CSMA and CSMA/CA fare better in both overall throughput and fairness because these protocols carrier senses before transmitting both TCP data and TCP ACKs. A busy channel triggers exponential backoff and thus allows one of the nodes to transmit successfully. To further analyze the effects of PSMA, CSMA and CSMA/CA, we next consider the ring experiments (table 4). Under the ring experiments, PSMA achieves better fairness and aggregate throughput compare to that of CSMA and CSMA/CA. PSMA obtains an aggregate throughput of 3 Mbps while CSMA and CSMA/CA both achieve 2.1 Mbps. PSMA s throughput of 3 Mbps is quite good considering the fact that the maximum theoretical throughput achievable on the ring is 4 Mbps (i.e., two simultaneous transmissions at least 3 hops apart). We note here again that CSMA/CA has no benefit compare to CSMA. Without control frames to determine the status of the receiver, the collision avoidance mechanism of CSMA/CA does not work well and, thus, behaves similar to plain CSMA. Therefore, the behaviors of both protocols are comparable throughout all experiments. Finally, we examine multiple TCP flows sharing a common bottleneck. The performance of PSMA, CSMA and CSMA/CA is shown in tables 5 7. In all three scenarios, the protocols perform relatively the same, with no one protocol in particular standing out. As the number of hops on the bottleneck link increases, more connections are unable to get data through. With so much contention and possibility of interference between TCP data and TCP ACK packets and no control frames to coordinate the access of the channel, the Table 7 Bottleneck experiments, two hops, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA results are not surprising. Capture is evident throughout, no matter the protocol and no matter the scenario. From the above experiments, we see that CSMA/CA has no beneficial effect without the feedback given by control frames such as RTS/CTS or ACK exchanges. Thus, CSMA is the better alternative to CSMA/CA due to lesser overhead. Between PSMA and CSMA, it is a toss up. PSMA is superior in both overall throughput and fairness under the ring experiment, but looses to CSMA in the exposed terminal environment in terms of being fair. We need to note that although PSMA does not sense carrier before transmitting, when used under TCP the TCP ACKs provide feedback on the state of the destination node and thus emulate some form of carrier sensing RTS/CTS To isolate the impact of RTS/CTS control frames, we contrast the following protocols: PSMA/RTS/CTS, CSMA/ RTS/CTS and CSMA/CA/RTS/CTS to PSMA, CSMA and CSMA/CA. Furthermore, we also evaluate PSMA/RTS/ CTS/ACK, CSMA/RTS/CTS/ACK and CSMA/CA/RTS/ CTS/ACK against PSMA/ACK, CSMA/ACK and CSMA/ CA/ACK. Thus, we are merely affixing RTS/CTS control frames to PSMA, CSMA, CSMA/CA, PSMA/ACK, CSMA/ACK and CSMA/CA/ACK. In the variable number of hops (tables 8 and 9), we see that adding RTS/CTS control frames only increase the overhead since there is no channel access contention by competing sources. In fact, the overhead results in a loss of about

6 322 K. TANG ET AL. Table 8 Variable number of hops experiment, without ACKs, throughput in bits per second. Number of hops PSMA CSMA CSMA/CA PSMA/RTS/CTS CSMA/RTS/CTS CSMA/CA/RTS/CTS Table 9 Variable number of hops experiment, with ACKs, throughput in bits per second. Number of PSMA/ CSMA/ CSMA/CA/ PSMA/RTS/CTS/ CSMA/RTS/CTS/ CSMA/CA/RTS/CTS/ hops ACK ACK ACK ACK ACK ACK Table 10 Hidden terminal experiment, without ACKs, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA PSMA/RTS/CTS CSMA/RTS/CTS CSMA/CA/RTS/CTS Table 11 Hidden terminal experiment, with ACKs, throughput in bits per second. TCP connection PSMA/ CSMA/ CSMA/CA/ PSMA/RTS/CTS/ CSMA/RTS/CTS/ CSMA/CA/RTS/CTS/ ACK ACK ACK ACK ACK ACK Table 12 Exposed terminal experiment, without ACKs, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA PSMA/RTS/CTS CSMA/RTS/CTS CSMA/CA/RTS/CTS Table 13 Exposed terminal experiment, with ACKs, throughput in bits per second. TCP PSMA/ CSMA/ CSMA/CA/ PSMA/RTS/CTS/ CSMA/RTS/CTS/ CSMA/CA/RTS/CTS/ connection ACK ACK ACK ACK ACK ACK Kbps on average, although the overhead cost becomes minimal as we increase the number of hops. We now look at the hidden terminal experiments (tables 10 and 11). Using RTS/CTS with PSMA provides no benefit. However, when combined with CSMA and CSMA/CA, fairness is achieved. The RTS/CTS exchange allows the competing nodes to be aware of each other and thus they are able to better coordinate for equal channel access. With CSMA/RTS/CTS, fairness is achieved without sacrificing the overall throughput of the network. Aggregate throughput is lower with the introduction of RTS/CTS to CSMA/CA. Adding RTS/CTS to PSMA/ACK, CSMA/ ACK and CSMA/CA/ACK also aids in fairness. Merging RTS/CTS with PSMA/ACK and CSMA/CA/ACK results in excellent fair sharing between the connection from node 0 to node 1 and node 2 to node 1. Combining RTS/CTS to CSMA/ACK has little effect. For the exposed terminal experiments (tables 12 and 13), the addition of the RTS/CTS to PSMA, CSMA and CSMA/ CA has a major impact on PSMA but does little for CSMA

7 EFFECTS OF AD HOC MAC LAYER MEDIUM ACCESS MECHANISMS 323 Table 14 Ring experiments, without ACKs, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA PSMA/RTS/CTS CSMA/RTS/CTS CSMA/CA/RTS/CTS Table 15 Ring experiments, with ACKs, throughput in bits per second. TCP PSMA/ CSMA/ CSMA/CA/ PSMA/RTS/CTS/ACK CSMA/RTS/CTS/ACK CSMA/CA/RTS/CTS/ connection ACK ACK ACK ACK Table 16 Bottleneck experiment, without ACKs, zero hops, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA PSMA/RTS/CTS CSMA/RTS/CTS CSMA/CA/RTS/CTS and CSMA/CA. PSMA/RTS/CTS alleviates most of the capture characteristics of PSMA and at the same time maintains good aggregate throughput. The RTS/CTS acts like a virtual carrier sense for PSMA, and thus, we expect the performance of PSMA/RTS/CTS to be similar to that of CSMA. This is consistent with the hidden terminal experiments where the virtual carrier sense emulation of the RTS/CTS control frames of PSMA/RTS/CTS behaves like CSMA and thus exhibit similar network behavior as that of CSMA with the hidden terminal environment. When incorporating RTS/ CTS with PSMA/ACK, CSMA/ACK and CSMA/CA/ACK, the RTS/CTS control frames again provides better fairness with PSMA/ACK. With just PSMA/ACK, the connection between node 2 to node 3 locks out the connection from node 1 to node 0, achieving a throughput of 1.76 Mbps. By introducing RTS/CTS, both sessions are able to more equally share the channel, with the node 2 to node 3 session getting 687 Kbps while the node 1 to node 0 achieving 988 Kbps. As for CSMA/ACK and CSMA/CA/ACK, adding RTS/CTS results in comparable network performances. As we move to the ring experiments, we see that total throughput increases with the use of RTS/CTS on top of PSMA, CSMA and CSMA/CA (tables 14 and 15). Aggregate throughputs of PSMA, CSMA and CSMA/CA in the ring experiment are 3.03, 2.08 and 2.06 Mbps, respectively, comparing to that of PSMA/RTS/CTS, CSMA/RTS/CTS and CSMA/CA/RTS/CTS, whose throughputs are 3.17, 3.34 and 2.70 Mbps, respectively. Using RTS/CTS with PSMA renders the system less fair while the opposite is true with CSMA and CSMA/CA where the amount of fair sharing is increased. However, when combining RTS/CTS with PSMA/ACK, CSMA/ACK and CSMA/CA/ACK, we observe that the RTS/CTS actually degrades the fair sharing of the system. Thus, too much control frames can be detrimental to the fairing sharing of the wireless channel. Throughput of PSMA/ACK ranges from 446 to 238 Kbps while PSMA/RTS/CTS/ACK ranges from 845 to 79 Kbps. With CSMA/ACK, minimum and maximum throughput is 444 and 275 Kbps, respectively, while CSMA/ACK/RTS/ CTS exhibited minimum and maximum throughput of 804 and 37 Kbps, respectively. Similarly, CSMA/CA/ACK shows throughput ranging from 555 to 204 Kbps while CSMA/CA/ACK/RTS/CTS shows throughput from 566 to 167 Kbps. Under the bottleneck experiments (tables 16 21), the addition of RTS/CTS to PSMA, CSMA and CSMA/CA in-

8 324 K. TANG ET AL. Table 17 Bottleneck experiment, with ACKs, zero hops, throughput in bits per second. TCP PSMA/ CSMA/ CSMA/CA/ PSMA/RTS/CTS/ CSMA/RTS/CTS/ CSMA/CA/RTS/CTS/ connection ACK ACK ACK ACK ACK ACK Table 18 Bottleneck experiment, without ACKs, one hop, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA PSMA/RTS/CTS CSMA/RTS/CTS CSMA/CA/RTS/CTS Table 19 Bottleneck experiment, with ACKs, one hop, throughput in bits per second. TCP PSMA/ CSMA/ CSMA/CA/ PSMA/RTS/CTS/ CSMA/RTS/CTS/ CSMA/CA/RTS/CTS/ connection ACK ACK ACK ACK ACK ACK Table 20 Bottleneck experiment, without ACKs, two hops, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA PSMA/RTS/CTS CSMA/RTS/CTS CSMA/CA/RTS/CTS Table 21 Bottleneck experiment, with ACKs, two hops, throughput in bits per second. TCP PSMA/ CSMA/ CSMA/CA/ PSMA/RTS/CTS/ CSMA/RTS/CTS/ CSMA/CA/RTS/CTS/ connection ACK ACK ACK ACK ACK ACK creases the collective throughput of the system under the various bottleneck scenarios. The degree of fairness is also increased. We observe, however, that fairness begins to fade as the number of hops of the bottleneck link increases. When RTS/CTS is added to PSMA/ACK, CSMA/ACK and CSMA/CA/ACK, the aggregate throughput again increases in all bottleneck scenarios. However, its effect on fairness is not quite universal. Using RTS/CTS with CSMA/ACK actually worsens fairness in the zero-hop bottleneck experiments. On the other hand, when included with CSMA/ CA/ACK or PSMA/ACK in the zero-hop bottleneck experiments, the RTS/CTS improves fairness. In the other bottleneck experiments, the RTS/CTS has no major impact in terms of fairness. Note that with CSMA/CA/ACK and CSMA/CA/ACK/RTS/CTS, fairness is preserved no matter the number of hops. From the above experiments, we conclude that for the most part, the RTS/CTS control frames do improve the fair-

9 EFFECTS OF AD HOC MAC LAYER MEDIUM ACCESS MECHANISMS 325 Table 22 Variable number of hops experiment, without RTS/CTS, throughput in bits per second. Number of hops PSMA CSMA CSMA/CA PSMA/ACK CSMA/ACK CSMA/CA/ACK Table 23 Variable number of hops experiment, with RTS/CTS, throughput in bits per second. Number of PSMA/ CSMA/ CSMA/CA/ PSMA/ CSMA/ CSMA/CA/ hops RTS/CTS RTS/CTS RTS/CTS ACK ACK ACK Table 24 Hidden terminal experiment, without RTS/CTS, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA PSMA/ACK CSMA/ACK CSMA/CA/ACK Table 25 Hidden terminal experiment, with RTS/CTS, throughput in bits per second. TCP PSMA/ CSMA/ CSMA/CA/ PSMA/ CSMA/ CSMA/CA/ connection RTS/CTS RTS/CTS RTS/CTS RTS/CTS/ACK RTS/CTS/ACK RTS/CTS/ACK ness characteristics of the network and in some cases also increase the aggregate throughput. However, there are times when the combination of RTS/CTS and ACK control frames will not help but instead hurt the fair sharing of the channel ACK To determine the effects of ACKs, the following protocols are put in juxtaposition: PSMA/ACK, CSMA/ACK and CSMA/CA/ACK to PSMA, CSMA and CSMA/CA. In addition, we also compare PSMA/RTS/CTS/ACK, CSMA/RTS/ CTS/ACK and CSMA/CA/RTS/CTS/ACK to PSMA/RTS/ CTS, CSMA/RTS/CTS and CSMA/CA/RTS/CTS. We start again with the variable number of hops experiment. From tables 22 and 23, we see as before that the ACKs only adds overhead here since there is no channel contention. The overhead for PSMA/ACK, CMSA/ACK, PSMA/RTS/CTS/ACK and CSMA/RTS/CTS/ACK is approximately 70 Kbps compared to that of PSMA, CSMA, PSMA/RTS/CTS and CSMA/RTS/CTS, respectively. However, the overhead for using ACK frames in CSMA/CA/ ACK and CSMA/CS/RTS/CTS/ACK is more than 200 Kbps. Still, as the number of hops increases, the overhead is reduced substantially. Examining the results of the hidden terminal experiments (tables 24 and 25), we learn that ACKs give mixed results regarding the fair sharing of the network. Adding ACKs to PSMA had little effect while adding ACKs to CSMA and CSMA/CA increases fairness, although not totally. When used with PSMA/RTS/CTS, ACKs greatly level the throughput of both connections. However, putting ACKs with CSMA/RTS/CTS actually worsens fair sharing. Using ACKs with CSMA/CA/RTS/CTS added no benefit. In terms of throughput, when combined with PSMA (1.74 vs Mbps), PSMA/RTS/CTS (1.74 vs Mbps), CSMA (1.75 vs Mbps) and CSMA/RTS/CTS ( vs bps), link-level ACKs improve the aggregate throughput while the throughput deteriorates with CSMA/ CA (1.72 vs Mbps) and CSMA/CA/RTS/CTS (1.61 vs Mbps). In the exposed terminal experiments, link-level ACKs demonstrate no improvements in either fairness or total throughput. As a matter of fact, aggregate throughput decreases slightly (tables 26 and 27). The lack of improvement when ACKs are introduced is attributed to the fact that

10 326 K. TANG ET AL. Table 26 Exposed terminal experiment, without RTS/CTS, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA PSMA/ACK CSMA/ACK CSMA/CA/ACK Table 27 Exposed terminal experiment, with RTS/CTS, throughput in bits per second. TCP connection PSMA/ CSMA/ CSMA/CA/ PSMA/ CSMA/ CSMA/CA/ RTS/CTS RTS/CTS RTS/CTS RTS/CTS/ACK RTS/CTS/ACK RTS/CTS/ACK Table 28 Ring experiment, without RTS/CTS, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA PSMA/ACK CSMA/ACK CSMA/CA/ACK Table 29 Ring experiment, with RTS/CTS, throughput in bits per second. TCP PSMA/ CSMA/ CSMA/CA/ PSMA/ CSMA/ CSMA/CA/ connection RTS/CTS RTS/CTS RTS/CTS RTS/CTS/ACK RTS/CTS/ACK RTS/CTS/ACK Table 30 Bottleneck experiment, without RTS/CTS, zero hops, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA PSMA/ACK CSMA/ACK CSMA/CA/ACK in the exposed terminal scenario, link-level ACKs in most cases will not experience any collisions and, therefore, will not greatly affect the overall performance of the network in the exposed terminal environment. Using RTS/CTS control frames by themselves is sufficient here in the exposed terminal environment. Moving on to the ring experiments (tables 28 and 29), we see that ACKs with PSMA had little effect. However, with combined with CSMA or CSMA/CA, the overall throughput dramatically increases, from 2.1 Mbps in CSMA to 3.2 Mbps in CSMA/ACK and from 2.1 Mbps in CSMA/CA to 2.7 Mbps in CSMA/CA/ACK. The same pattern occurs when we introduce ACKs to PSMA/RTS/CTS, CSMA/RTS/CTS and CSMA/CA/RTS/CTS, although to a lesser extent. It is uncertain whether or not ACKs have any beneficial outcomes in terms of fairness under the ring experiment. Under the bottleneck experiments (tables 30 35), the addition of ACKs again is not beneficial when combined with PSMA/ACK and CSMA/ACK. And once more, us-

11 EFFECTS OF AD HOC MAC LAYER MEDIUM ACCESS MECHANISMS 327 Table 31 Bottleneck experiment, with RTS/CTS, zero hops, throughput in bits per second. TCP PSMA/ CSMA/ CSMA/CA/ PSMA/ CSMA/ CSMA/CA/ connection RTS/CTS RTS/CTS RTS/CTS RTS/CTS/ACK RTS/CTS/ACK RTS/CTS/ACK Table 32 Bottleneck experiment, without RTS/CTS, one hop, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA PSMA/ACK CSMA/ACK CSMA/CA/ACK Table 33 Bottleneck experiment, with RTS/CTS, one hop, throughput in bits per second. TCP PSMA/ CSMA/ CSMA/CA/ PSMA/ CSMA/ CSMA/CA/ connection RTS/CTS RTS/CTS RTS/CTS RTS/CTS/ACK RTS/CTS/ACK RTS/CTS/ACK Table 34 Bottleneck experiment, without RTS/CTS, two hops, throughput in bits per second. TCP connection PSMA CSMA CSMA/CA PSMA/ACK CSMA/ACK CSMA/CA/ACK Table 35 Bottleneck experiment, with RTS/CTS, two hops, throughput in bits per second. TCP PSMA/ CSMA/ CSMA/CA/ PSMA/ CSMA/ CSMA/CA/ connection RTS/CTS RTS/CTS RTS/CTS RTS/CTS/ACK RTS/CTS/ACK RTS/CTS/ACK ing ACKs with CSMA/CA shows improvement over simply using CSMA/CA due the feedback that occurs when ACK frames are lost. From the results of the experiments above, the conclusion is that ACKs in general improves the combined throughput of the network but its effect on fairness is mixed Summary From the observations in section 6.1, the resolve is that PSMA and CSMA works best overall compared to CSMA/ CA in terms of overall network throughput and fairness. By examining section 6.2, we see that RTS/CTS control frames

12 328 K. TANG ET AL. assist to provide fairness to the network and in most circumstances improves the aggregate throughput of the network as well. We also discover that although RTS/CTS control frames when appended to PSMA/ACK, CSMA/ACK and CSMA/CA/ACK in general help reduce capture effects, there are certain circumstances that too much control frames are deleterious to fair sharing. Finally, section 6.3 leads us deduce that ACKs in general improve the cumulative throughput of the network environment under study. The combination that works best together is CSMA/CA/ RTS/CTS/ACK when running FTP. CSMA/CA/RTS/CTS/ ACK provides a superior mixture of fairness and aggregate network throughput under the seven topologies being studied. What distinguishes CSMA/CA/RTS/CTS/ACK from the rest of the protocols we analyze is the fact that even when competing traffics are several hops away from one another, CSMA/CA/RTS/CTS/ACK still manages to provide some sort of channel access coordination among the competing streams. 7. Conclusion From the UDP and TCP experiments that we conducted, the general trend is that carrier sensing by itself is superior to packet sensing and carrier sensing with collision avoidance. The RTS/CTS mechanism is mainly responsible for channel access coordination, which in turn leads to greater fair sharing of the network resource and in some cases improves total throughput. With ACK control frames, more often than not, aggregate throughput is increased due to the persistent retransmission when ACK frames are lost. However, sometimes fairness is sacrificed. The combination of CSMA/RTS/CTS/ACK is the favorite for UDP traffic with CBR, although CSMA/CA/RTS/ CTS/ACK comes in a close second. On the TCP side running FTP, CSMA/CA/RTS/CTS/ACK is the clear winner. Thus, we conclude that CSMA/CA/RTS/CTS/ACK provides the best overall network service under general terms. Not surprising, CSMA/CA/RTS/CTS/ACK is exactly the IEEE standard with virtual carrier sense enabled. Two other important mechanisms should also be considered. Specifically, the backoff schemes and the yield time used in MAC protocols. Yield time is defined as the amount of time the sender backs off before sending another data frame after transmitting a data frame. Yield time also refers to the amount of time the receiver backs off before sending a frame after receiving a data frame. By adjusting the backoff schemes and yield time parameters, we can conceivably manipulate the outcome of the network performance. For example, if we implement a less aggressive backoff timer or use a longer yield time, fairness can be achieved. Therefore, further research is needed regarding the MAC layer timers and their effects on the protocols studied in this paper. References [1] R. Bagrodia, R. Meyerr et al., PARSEC: A parallel simulation environment for complex system, Computer Magazine (1998). [2] V. Bharghavan, A. Demers, S. Shenker and L. Zhang, MACAW: A media access protocol for wireless LAN s, in: ACM SIGCOMM (1994). [3] M. Correa and K. Tang, Isolation of wireless ad hoc medium access mechanisms under UDP, (May 28, 1999). [4] C. Fullmer and J.J. Garcia-Luna-Aceves, Floor acquisition multiple access (FAMA) for packet radio networks, Computer Communication Review 25(4) (October 1995). Also in: ACM SIGCOMM 95, Cambridge, MA, USA (28 August 1 September 1995). [5] M. Gerla, K. Tang and R. Bagrodia, TCP performance in wireless multihop networks, in: Proceedings of IEEE WMCSA 99 (February 1999). [6] J. Haartsen, M. Naghshineh, J. Inouye, O.J. Joeressen and W. Allen, Bluetooth: Vision, goals, and architecture, ACM SIGMOBILE Mobile Computing and Communications Review 2(4) (October 1998) [7] IEEE , Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications, Draft standard, P802.11/D5.0 (July 1996). [8] A. Joseph, B.R. Badrinath and R. Katz, A case for services over cascaded networks, in: First ACM/IEEE International Conference on Wireless and Mobile Multimedia (WoWMoM 98), Dallas, TX (October 1998). [9] J. Jubin and J.D. Tornow, The DARPA packet radio network protocols, Proceedings of the IEEE (January 1987). [10] P. Karn, MACA a new channel access method for packet radio, in: ARRL/CRRL Amateur Radio 9th Computer Networking Conference, ARRL (1990). [11] K.J. Negus, J. Waters, J. Tourrilhes, C. Romans, J. Lansford and S. Hui, HomeRF and SWAP: Wireless networking for the connected home, ACM SIGMOBILE Mobile Computing and Communications Review 2(4) (October 1998) [12] A.S. Tanenbaum, Computer Networks, 3rd ed. (Prentice Hall, Englewood Cliffs, NJ, 1996). [13] X. Zeng, R. Bagrodia and M. Gerla, GloMoSim: a library for the parallel simulation of large-scale wireless networks, PADS (1998). Ken Tang graduated from the University of California, Irvine (UCI), with a Bachelor of Science degree in information and computer science in He then received his Masters of Science degree in computer science in 1999 from the University of California, Los Angeles (UCLA). Currently, Ken is a Ph.D. student at UCLA. His research interest is in the area of ad hoc networks, specifically focusing on ad hoc medium access control protocols. Ken is a member of ACM. ktang@cs.ucla.edu Mario Correa graduated from California State University, Los Angeles, with a Bachelor degree in mathematics in He is currently a graduate student in computer science at the University of California, Los Angeles (UCLA). His research interest is in the area of wireless networks. mcorrea@cs.ucla.edu Mario Gerla received his graduate degree in electrical engineering from Politecnico di Milano, Italy, in 1966 and his M.S. and Ph.D. degrees in computer science from University of California, Los Angeles, in 1970 and 1973, respectively. From 1973 to 1976, Dr. Gerla was a manager in Network Analysis Corporation, Glen Cove, NY, where he was involved in several computer network design projects for both government and industry, including performance analysis and topological updating of the ARPANET under a contract from DoD. From 1976 to 1977, he was with Tran Telecommunication, Los Angeles, CA, where he participated in the development of an integrated packet and circuit network. Since 1977, he

13 EFFECTS OF AD HOC MAC LAYER MEDIUM ACCESS MECHANISMS 329 has been on the Faculty of the Computer Science Department of UCLA. His research interests include the design, performance evaluation, and control of distributed computer communication systems and networks. His current research projects cover the following areas: design and performance evaluation of protocols and control schemes for ad hoc wireless networks; routing, multicasting, congestion control and bandwidth allocation in wide area networks, and traffic measurements and characterization. Dr. Gerla currently serves on the editorial board for IEEE/ACM Transactions on Networking and ACM/Kluwer Journal of Wireless Networks (WINE).