258
IEEE TRANSACTIONS ON MOBILE COMPUTING,
VOL. 5,
NO. 3,
MARCH 2006
A Wireless MAC Protocol Using Implicit Pipelining Xue Yang, Member, IEEE, and Nitin H. Vaidya, Senior Member, IEEE Abstract—In distributed multiple access control protocols, two categories of overhead are usually associated with contention resolution. One is channel idle overhead, where all contending stations are waiting to transmit. Another is collision overhead, which occurs when multiple contending stations attempt to transmit simultaneously. Either idle overhead or collision overhead being large, contention resolution algorithm would be inefficient. Prior research work tries to minimize both the idle and the collision overheads using various methods. In this paper, we propose to apply “pipelining” techniques to the design of multiple access control protocol so that channel idle overhead could be (partially) hidden and the collision overhead could be reduced. While the concept of pipelined scheduling can be applied to various MAC protocol designs in general, in this paper, we focus on its application to IEEE 802.11 DCF. In particular, an implicitly pipelined dual-stage contention resolution MAC protocol (named DSCR) is proposed. With IEEE 802.11, the efficiency of contention resolution degrades dramatically with the increasing load due to high probability of collision. Using the implicit pipelining technique, DSCR hides the majority of channel idle time and reduces the collision probability, hence, improves channel utilization, average access delay, and access energy cost over 802.11 significantly both in wireless LANs and in multihop networks. The simulation results, as well as some analysis, are presented to demonstrate the effectiveness of DSCR. Index Terms—Multiple access control (MAC), IEEE 802.11, pipelining, wireless LANs, multihop networks, channel utilization, access energy cost, packet access delay.
æ 1
I
INTRODUCTION
802.11 standard [1] defines a distributed coordination function (DCF), which uses a binary exponential backoff (BEB) algorithm to resolve channel contention. BEB algorithm controls the channel contention by adjusting the value of CW (CW represents contention window and is a parameter of 802.11). When a station successfully transmits a packet, it resets its CW to a minimum value CWmin ; when a collision occurs, the colliding stations exponentially increase their CW by a factor of 2, until CW reaches the maximum value CWmax . A station wanting to access the channel generates a random backoff counter uniformly distributed over the interval [0, CW]. This backoff counter corresponds to the number of idle slots this station has to wait before its transmission attempt. Clearly, the choice of CW is critical to the performance of 802.11. When there are few contending stations in the network, a smaller CW will reduce the channel idle time and enable a better usage of channel bandwidth. When the number of contending stations increases, a larger CW is preferred to reduce the collision probability. Some prior research analyzes the performance of IEEE 802.11 DCF [2], [3], [4], [5], [6], [7], [8], [9], and shows that IEEE 802.11 DCF operates far from the optimal points of CW, which can be confirmed from the simulation results shown in Fig. 1 and other papers as well [2], [5], [10]. The simulation results in EEE
. The authors are with the Department of Electrical and Computer Engineering and the Coordinated Science Laboratory, University of Illinois at Urbana-Champaign, 1308 West Main Street, Urbana, IL 61801. E-mail: {xueyang, nhv}@uiuc.edu. Manuscript received 24 Sept. 2003; revised 12 Aug. 2004; accepted 8 Oct. 2004; published online 16 Jan. 2006. For information on obtaining reprints of this article, please send e-mail to:
[email protected], and reference IEEECS Log Number TMC-0155-0903. 1536-1233/06/$20.00 ß 2006 IEEE
Fig. 1 are obtained using ns-2 network simulator for wireless LANs, in which all stations can hear each other’s transmission. The payload size is 512 bytes and Constant Bit Rate traffic is used. We can see that the peak throughput of 802.11 DCF is achieved when there are four contending stations, and the throughput degrades with fewer or greater number of contending stations. With less than four contending stations, the unnecessary channel idle time is the primary reason for the throughput degradation. On the other hand, when increasing the number of contending stations from 4 to 256, the aggregate throughput of 802.11 DCF degrades from 91 percent of ideal throughput (defined as the maximum throughput a network can obtain without any MAC scheduling idle and collision overhead) to 68 percent, which is mainly due to the increased packet collisions. Two categories of overhead are associated with contention resolution at MAC layer. One is channel idle overhead, where all stations are waiting to transmit. The other is collision overhead, which occurs when multiple contending stations attempt to transmit simultaneously. In 802.11 DCF, stations perform various operations sequentially. Contending stations go through contention resolution procedure to determine which station has the right of accessing the channel; then, the winning station transmits its packet. Only when current transmission finishes, a new round of contention resolution begins—for ease of exposition, assume a LAN scenario here. In the sequential procedure, both channel idle time and collisions consume the entire channel bandwidth, resulting in performance tradeoffs. If a larger CW is applied, 802.11 DCF can achieve better throughput in heavily contended networks due to the reduced collision probability. However, at the same time, the channel idle overhead in networks with little contention Published by the IEEE CS, CASS, ComSoc, IES, & SPS
YANG AND H. VAIDYA: A WIRELESS MAC PROTOCOL USING IMPLICIT PIPELINING
259
MAC protocol, named DSCR, in detail. Section 5 presents the simulation results of DSCR for wireless LANs. The improved contention resolution using DSCR in multihop networks is discussed in Section 6. Related work is discussed in Section 7. Conclusions are drawn in Section 8. Finally, theoretical analysis regarding the reduced channel contention using DSCR is given in the Appendix which is provided for free at http://computer.org/tmc/archives.htm.
2
Fig. 1. Throughput of 802.11 versus the number of contending stations (throughput is normalized to the ideal value with perfect scheduling).
increases using a larger CW, which leads to the degraded throughput in such networks. Traditionally, efforts have been made to dynamically adjust stations’ channel access behavior based on the network contention status so that both the channel idle and collision overhead can be minimized [10], [11], [12], [13], [14]. However, such algorithms typically require extensive channel feedback information, which may not be available in wireless networks, to infer the network contention status. We consider an alternative method, which applies “pipelining” techniques to MAC protocol design. Pipelining has been used successfully in other areas (e.g., computer architecture [15]) to improve performance. The key mechanism used for pipelining is to divide the total task into subtasks, and to introduce parallelism by allowing different subtasks of different tasks proceed simultaneously. For MAC protocols, the total task is to schedule the channel access and transmit the packet. By using “pipelining,” the contention resolution procedure that schedules the channel access (partially) overlaps with the packet transmission duration. That is, when a pair of source and destination stations are using the channel to exchange packets, the remaining contending stations start the contention resolution procedure in parallel in order to resolve the channel contention for next packet. Since it is performed in parallel with packet transmission, the pipelined contention resolution procedure consumes little channel bandwidth, which can help to improve performance. Our previous research has explored explicit pipelining, wherein a control channel is used for explicitly performing pipelined contention resolution [16], [17]. This paper considers an implicit pipelining scheme that attempts to retain the benefits of explicit pipelining, but without using a separate control channel. The performance evaluation results presented later in this paper indicate that the implicit pipelining scheme is able to achieve good performance. The rest of the paper is organized as follows: A brief introduction to IEEE 802.11 DCF is presented in Section 2; the proposed scheme bears significant similarities to IEEE 802.11. In Section 3, one explicit pipelining scheme is discussed, which motivates the proposed “implicit pipelining” mechanism. Section 4 describes the proposed implicit pipelining
OVERVIEW
OF
IEEE 802.11 DCF
We briefly describe the relevant features of IEEE 802.11 DCF here. For more details, please refer to [1]. IEEE 802.11 DCF defines two access methods: basic access method and RTS/ CTS access method. The basic access method involves only Data/ACK exchange, in which data packets are transmitted when channel access has been obtained. ACK frames follow successful data packet receptions. In the RTS/CTS access method, RTS (Request To Send) and CTS (Clear To Send) frames are exchanged before Data/ACK packets. RTS and CTS frames contain a duration field that defines the period of time for which the medium is to be reserved to transmit the actual Data frame and the returning ACK frame. Stations which overhear RTS/CTS frames defer transmission for this period. This mechanism is referred to as “virtual carrier sensing” and it is implemented using “Network Allocation Vector” (NAV). The duration field is also available in the MAC header of Data and ACK frames. A station updates the NAV with the duration field specified in the overheard frames. The carrier sense mechanism in IEEE 802.11 includes physical carrier sense and virtual carrier sense. After channel is sensed idle for a DIFS (DCF Interframe Space) duration, the backoff procedure is invoked by a backlogged station and its backoff counter is decremented by 1 after each idle slot. As we explained in Section 1, 802.11 DCF uses binary exponential backoff (BEB) algorithm to resolve channel contention. A shorter interframe space, SIFS, is used to separate transmissions belonging to a single dialog (e.g., CTS, Data, and ACK frames in the case of RTS/CTS access method). Fig. 2 illustrates the RTS/CTS access method of IEEE 802.11 DCF. In terms of channel feedback, a station using 802.11 receives feedback from its own packet transmission (success or failure) to adjust its contention window size. In addition, the physical and virtual carrier sensing help to avoid potential collisions.
3
PIPELINED PACKET SCHEDULING
To perform pipelined packet scheduling, the pipelined contention resolution procedure can either be used to completely resolve the channel contention or only partially resolve the contention [16], [17], [18]. When it is used to resolve the channel contention partially, the purpose of the pipelined contention resolution procedure is to hide a part of channel idle overhead and to reduce collision overhead. In our prior work, we have developed a “partial pipelining” scheme [16], [17], which is briefly discussed below to motivate the proposed implicit pipelining scheme. The
260
IEEE TRANSACTIONS ON MOBILE COMPUTING,
VOL. 5,
NO. 3,
MARCH 2006
Fig. 2. RTS/CTS access method of IEEE 802.11 DCF.
Fig. 3. Partial pipelining scheme.
partial pipelining scheme uses a narrow-band busy tone channel for the purpose of pipelining, as explained next. In partial pipelining, the contention resolution procedure is split into two phases, as shown in Fig. 3. Pipelined stage 1 includes only contention resolution phase 1 and is performed on the busy tone channel. Contention resolution phase 2 and packet transmissions are performed on the data channel in pipelined stage 2. A backoff counter is associated with contention resolution phase 1. When some other station is transmitting on the data channel, a contending station reduces its phase 1 backoff counter by 1 after each idle slot on the busy tone channel. Whenever the phase 1 backoff counter reaches zero, the contending station sends out a signal on the busy tone channel to claim its winning status and becomes a “pipelined station.” In the rest of this paper, we use the term “pipelined station” to refer to a station that has won contention resolution phase 1 and is allowed to enter contention resolution phase 2, while the term “contending station” is used for any backlogged station that wants to access the channel. Only pipelined stations are allowed to contend for the data channel access when current transmission finishes. Other contending stations freeze themselves in contention resolution phase 1 upon sensing a busy tone signal. The data channel contention among pipelined stations is further resolved in contention resolution phase 2 following the procedure similar to 802.11 DCF (but with smaller contention window sizes). It can be shown that, with a large probability, the number of pipelined stations is small with perfect busy tone detection. Thus, the collision probability among the pipelined stations is small, which leads to the reduced collision overhead. Also, since a significant portion of random backoff is performed on the
narrow-band busy tone channel in contention resolution phase 1 when data channel is busy, the channel idle overhead associated with the backoff procedure is effectively reduced. Partial pipelining can achieve a significant improvement over IEEE 802.11 in terms of channel utilization across a wide range of network sizes [16], [17]. However, partial pipelining relies on a busy tone channel, which incurs more hardware cost. Additionally, in wireless networks, the busy tone signal may not be sensed reliably. In fact, with hidden terminals, the busy tone transmitted by a station may not be sensed by another station at all. With unreliable busy tone detection, the performance of partial pipelining may be expected to degrade—to assess the performance impact of imperfect busy tone detection, we evaluated the partial pipelining scheme with perfect busy tone detection (i.e., 100 percent detection probability) and the worst-case scenario of no busy tone detection (0 percent detection probability). The performance results for wireless LANs with up to 256 contending stations are plotted in Fig. 4. As expected, the performance of partial pipelining degrades with 0 percent detection probability for busy tones. However, surprisingly, the performance remains superior to IEEE 802.11. A closer look at partial pipelining reveals reasons for this phenomenon. On the one hand, lack of busy tone detection implies that there is no way for a pipelined station to signal other stations to freeze themselves in contention resolution phase 1; this increases number of contending stations in stage 2, resulting in poorer performance compared with perfect busy tone detection. On the other hand, contention resolution phase 1 allows only a subset of stations to enter stage 2 at any given time,
YANG AND H. VAIDYA: A WIRELESS MAC PROTOCOL USING IMPLICIT PIPELINING
4
Fig. 4. Performance of partial pipelining with and without busy tone detection (packet size: 512 bytes).
resulting in better performance than IEEE 802.11. These observations motivated us to consider development of a MAC protocol that mimics partial pipelining, but without using a separate control channel. Effectively, our goal is to achieve performance of perfect busy tone detection without using a busy tone at all (note that 0 percent detection probability is equivalent to not transmitting a busy tone). Again, consider the partial pipelining scheme with 0 percent busy tone detection probability. In this case, when a station counts down its phase 1 counter to 0, and transmits a busy tone, other stations do not sense the busy tone, and can continue counting down their backoff counters. Effectively, stations can count down their phase 1 backoff counters for the duration of the on-going data packet transmission on the data channel—if the data packet transmission lasts for L slots, then stations will be able to count down their counters by L slots. Thus, partial pipelining with 0 percent busy tone detection can be implemented by simply allowing each station to decrement its phase 1 backoff counter by L slots whenever a data packet transmission is detected as having been completed (instead of decrementing the counter by 1 after each slot time). Next, instead of using L, the data packet duration, one may decrement the backoff counter by any suitable amount. The implicit pipelining illustrated in Fig. 5 is derived from the above observations and incorporates other adaptations of backoff procedures to improve performance. The proposed MAC protocol DSCR (pipelined DualStage Contention Resolution) is described in the next section in detail.
Fig. 5. Implicitly pipelined packet scheduling.
261
PIPELINED DUAL STAGE CONTENTION RESOLUTION MAC PROTOCOL (DSCR)
DSCR uses backoff mechanism analogous to 802.11 DCF. Each contention resolution phase maintains its own contention window and backoff counter. As in 802.11 DCF, a station using DSCR receives feedback from its own packet transmission (success or failure) to adjust its contention window size. Available physical and virtual carrier sense (using NAV mechanism) also help to avoid some potential collisions. DSCR differs from 802.11 DCF only in the contention resolution algorithm, while the other functions remain the same. For heavily contended networks, DSCR statistically controls the number of pipelined stations to be relatively small so that channel contention in stage 2 can be resolved efficiently. At the same time, DSCR is carefully designed to avoid unnecessary waste of channel bandwidth for networks with little contention, which will get clearer in view of protocol details.
4.1 Contention Resolution Phase 1 (Stage 1) CW1 is the contention window for contention resolution phase 1 and it has a minimum value CW 1min and a maximum value CW 1max . The initial value of CW1 is CW 1min . A backoff counter, named bc1 , is associated with contention resolution phase 1. bc1 is chosen to be uniformly distributed over the interval [0, CW1]. Whenever a station’s bc1 becomes less than or equal to 0, this station becomes a pipelined station and enters stage 2. Updating of bc1 : The backoff counter bc1 is reduced in the following two ways: 1.
At the end of an overheard successful packet transmission, a contending station reduces its bc1 by a quantity named F . By doing this, the pipelined contention resolution phase 1 is implicitly performed. With a large value of F , many stations could have their bc1 0 after reducing it by F ; hence, many stations could enter stage 2 to contend for the channel. On the other hand, with a smaller F , there would be fewer stations entering stage 2. In DSCR, we choose F to be a function of tc, where tc represents the number of successfully transmitted packets overheard by the contending station ever since the most recent time it enters stage 1. tc will be reset to 1 when a station enters stage 1 for a newly arriving packet or when returns from stage 2 to
262
IEEE TRANSACTIONS ON MOBILE COMPUTING,
stage 1, and tc will be increased by 1 after each subsequent overheard successful transmission. While there are various choices possible for F , we currently define F as F ¼ 2tc 1. Thus, the longer a station has stayed in the stage 1, the more aggressively it reduces its bc1 , hence, the larger probability of becoming a pipelined station. Once bc1 0, the station enters stage 2 to contend for the channel. 2. It is possible that, when current transmission of a station finishes, none of the surrounding contending stations have their bc1 less than or equal to zero. When such cases occur, no pipelined stations contend for the channel in stage 2, channel bandwidth will be unnecessarily wasted. To avoid this, in DSCR, bc1 is also linearly decreased when channel is idle, i.e., bc1 is reduced by 1 after each idle slot. Whenever a station’s bc1 reaches zero, it becomes a pipelined station and enters stage 2. Updating of CW1: Among pipelined stations that contend for the channel in stage 2, a station that eventually wins channel access transmits its packet, then resets CW1 to CW 1min , resets tc to 1, regenerates bc1 from the interval [0, CW1], and returns to stage 1. On the other hand, a pipelined station that loses channel contention in stage 2 will double its CW1,1 regenerate bc1 , reset tc to 1 and return to stage 1. We will illustrate stage 1 later using some examples.
4.2 Contention Resolution Phase 2 CW2 is contention window for contention resolution phase 2. It has a minimum value CW 2min and a maximum value CW 2max . Initial value of CW2: The initial value of CW2 is CW 2min for all stations entering stage 2. A backoff counter, named bc2 , is associated with the contention resolution phase 2. Whenever bc2 reaches zero, a transmission is allowed. Initial value of bc2 : Depending on how a pipelined station enters stage 2, there are two different choices for the initial value of bc2 . 1.
2.
When a packet transmission finishes, multiple pipelined stations in the neighborhood could occur and enter stage 2 at the same time since they all may have bc1 0 after reducing bc1 by F . To further resolve the channel contention among these multiple pipelined stations, each of them generates an initial value for bc2 , which is uniformly distributed over the interval [0, CW2]. As we mentioned in Section 4.1, a contending station will also reduce bc1 by 1 after each slot when the channel is sensed idle. When a station counts down its bc1 to zero in such a way, it will become a pipelined station and enter stage 2 with the initial value of bc2 being 0. This is due to the following considerations: The objective of contention resolution phase 2 is to further resolve the channel contention among multiple pipelined stations occurred at the same time. Since the probability that more than
1. When we say “double” CW1, we mean that CW1 = 2*CW1 + 1. For instance, if CW1 is originally 7, on doubling, it becomes 15.
VOL. 5,
NO. 3,
MARCH 2006
one station count down bc1 to 0 simultaneously during channel idle time is small, a further phase 2 backoff may lead to unnecessary channel waste. Transmissions and Retransmissions: Despite different initial values of bc2 the pipelined stations in stage 2 may have, they all follow the same rules for transmissions and retransmissions. A pipelined station has to wait for the channel to be idle for DIFS duration before the backoff procedure of contention resolution phase 2 begins (as in 802.11). Then, if a pipelined station’s bc2 reaches zero and the channel is idle, the station will begin its transmission; otherwise, bc2 will be decreased by 1 after each idle slot. Before bc2 of a pipelined station reaches zero, if a frame sent by some other station is overheard (e.g., RTS or CTS frames in the case of the RTS/CTS access method), the former pipelined station loses channel contention and returns to stage 1. When a collision2 happens, the colliding stations will double their CW2, and generate a new bc2 value from the interval [0, CW2]. The colliding stations as well as other pipelined stations all stay in stage 2 and repeat the above contention resolution phase 2 until someone eventually wins the channel. The pipelined station that wins the channel will transmit the packet, reset CW1 and CW2 to CW 1min and CW 2min , respectively, reset tc to 1, regenerate bc1 from the interval [0, CW1], and return back to the first stage. A pipelined station that loses the channel will double its CW1, reset CW2 to CW 2min , reset tc to 1, regenerate bc1 , and return back to the first stage.
4.3 Examples for DSCR Simulation results show that, for DSCR, CW 1min ¼ 15, CW 2min ¼ 31 are appropriate choices for networks up to 256 contending stations. The choices of CW 1max and CW 2max have no major impact on the performance of DSCR, provided that they are large enough to accommodate the maximum network size. We set both CW 1max and CW 2max to 1,023. Fig. 6 illustrates the dual stage contention resolution of DSCR. For ease of exposition, in the example, we pretend that all five stations observe the end of a packet transmission at time t0, and they all contend for the channel with each other. Notice it is possible that, in multihop networks, different stations observe different packet transmissions depending on their locations. We also pretend that all stations have tc ¼ 4 (i.e., four successfully transmitted packets have been overheard by the contending station since the most recent time it enters stage 1) at time t0, hence, each station reduces bc1 by 2tc 1 ¼ 15. Since stations 3 and 5 now have bc1 0, they enter stage 2 (becoming pipelined stations), generate bc2 from the interval [0, CW2] (the initial value of CW2 is CW 2min ¼ 31), and begin to count down bc2 .3 Stations 1, 2, and 4 stay in stage 1, reducing their bc1 by 1 after each idle slot. After seven slots, at time t2, station 5 counts down bc2 to zero and wins the channel access. Station 3 doubles its CW1 on losing stage 2, resets tc to 1 and returns back to stage 1. It also generates a new bc1 value 2. When an RTS is not followed by a CTS, or Data is not followed by an ACK, a collision is assumed to have occurred. 3. To simplify the explanation, we do not mention DIFS in this example.
YANG AND H. VAIDYA: A WIRELESS MAC PROTOCOL USING IMPLICIT PIPELINING
263
Fig. 6. An example for the dual stage contention resolution of DSCR (time axis is not drawn to scale).
Fig. 7. A single flow.
which is uniformly distributed over the interval [0, CW1]. At time t3 when station 5 finishes its transmission, station 5 resets CW1 to CW 1min ðCW 1min ¼ 15Þ, resets tc to 1, returns back to stage 1, and generates a new value for bc1 from the interval [0, CW1]. Upon station 5 finishing its transmission at time t3, all stations again reduce their bc1 by 2tc1 . Notice this time, stations 1, 2, and 4 have tc ¼ 5, hence, bc1 of these three stations is reduced by 31. On the other hand, stations 3 and 5 have tc ¼ 1 and their bc1 are reduced by 1. Now that all stations have their bc1 larger than 0, none of them enter stage 2 at time t3. However, bc1 of each station is also decreased by 1 after each idle slot, at time t4, bc1 of station 2 is counted down to zero after 5 idle slots. Station 2 then becomes a pipelined station, enters stage 2, sets the initial value of bc2 to 0 and transmits at time t4. A special case happens when there is only a single flow in the network. In the example of Fig. 7, at time t0, station 0 is the only backlogged source station in the network and its bc1 is 8. Assuming tc ¼ 1 at time t0, station 0 then has bc1 ¼ 7 after reducing bc1 by 2tc 1 ¼ 1, and it continues to stay in stage 1 because bc1 > 0. During channel idle time, bc1 is decreased by 1 after each slot. After seven idle slots, at
time t1, bc1 is counted down to zero. The initial value of bc2 is set to 0 and it begins to transmit its packet. At time t2 when station 0 finishes its transmission, it resets CW1 to CW 1min ðCW 1min ¼ 15Þ, resets tc to 1, and generates a new value, say 9, for bc1 . Again, at time t2, station 0 reduces bc1 by 2tc 1 ¼ 1. Then, after eight idle slots, the bc1 of station 0 is counted down to zero at time t3, and station 0 transmits again. As we can see, when there is only a single flow in the network, DSCR performs very similar to IEEE 802.11 DCF if CW 1min of DSCR has the same value as CWmin of 802.11.
4.4 Summary of DSCR DSCR uses two stages to resolve the channel contention, which is not a new idea. For instance, European Telecommunications Standards Institute (ETSI) High Performance European Radio LAN (HIPERLAN/1) MAC protocol [19] uses two stages, namely, the elimination stage and the yield stage, to resolve channel contention. In the elimination stage of HIPERLAN/1, a contending station transmits elimination burst for a random duration and then listens to the channel in the elimination survival verification interval. A contending station survives the
264
IEEE TRANSACTIONS ON MOBILE COMPUTING,
Fig. 8. Dynamic feedback of DSCR.
elimination stage if and only if the channel is sensed idle in its elimination survival verification interval; otherwise, this station is eliminated and withdraws from current channel competition. The yield stage then further resolves contention and reduces the number of stations allowed to transmit to 1. The fundamental difference between DSCR and conventional two-stage contention resolution algorithms is that, in DSCR, stage 1 proceeds in parallel with stage 2 using pipelining techniques, as illustrated in Fig. 5. When observing the channel activities, only stage 2, which includes the contention resolution phase 2 and packet transmissions, actually consumes the channel bandwidth. Stage 1 reduces both the channel idle and collision overhead without introducing any additional cost. On the other hand, in conventional two-stage contention resolution algorithms, both stages consume channel bandwidth. For example, in HIPERLAN/1, the elimination stage cannot effectively reduce the channel contention if it is too short. At the same time, since the elimination stage also consumes channel time, the longer the elimination stage, the more wasted channel bandwidth in elimination stage. As a result, the length of elimination stage of HIPERLAN/1 has to be carefully chosen to trade-off the corresponding channel wastage with the effect of reducing channel contention, which is very similar to the tradeoff of choosing CW value faced by 802.11 DCF, as mentioned at the beginning of this paper (Section 1). Furthermore, without requiring any additional signaling mechanism (e.g., burst signals or busy tones), DSCR exploits the dynamic feedback between two contention resolution stages to effectively control the number of pipelined stations in stage 2. As Fig. 8 illustrates, with a total of N contending stations at the input, stage 1 generates M pipelined stations. The value of M largely depends on the distribution of CW1 among all N contending stations. If all contending stations have large values for CW1, then bc1 among them tends to be widely distributed and only a small fraction of them will have their bc1 reaching zero at any point of time. Thus, only a small number of stations will become pipelined stations. More stations may become pipelined stations if all contending stations have smaller values for CW1. At the same time, stage 2 selects 1 winner from M pipelined stations to access the channel, and the remaining M 1 pipelined stations double their CW1. The output of stage 2 thus affects the distribution of CW1 among N contending stations (i.e., input of stage 1), which, in turn, affects the number of pipelined stations that may
VOL. 5,
NO. 3,
MARCH 2006
occur next time. Thus, despite the large variation of the total number of contending stations, the number of pipelined stations entering stage 2 remains within a small range. The interaction between the distribution of CW1 and the number of pipelined stations in stage 2 is captured by the analysis presented in the Appendix, which can be found on the Computer Society Digital Library at http://computer. org/tmc/archives.htm. Both the analysis and our simulation results show that DSCR successfully controls the channel contention in stage 2. With up to 256 contending stations, the average number of pipelined stations contending for the channel in stage 2 is less than 28.
5
PERFORMANCE EVALUATION WIRELESS LANS
OF
DSCR
IN
As wireless LAN environment gives us a simple context to reveal some key properties of DSCR, in this section, we use some simulation results from wireless LANs to continue our discussions on DSCR protocol. Extensive simulation results for multihop wireless networks will be presented in Section 6 to show the performance improvement of DSCR over 802.11 in multihop networks. All the simulation results in this paper are based on a modified version of ns-2 network simulator [20]. The effective transmission range is limited to 250 meters and carrier sense range (the range in which carrier can be sensed) is limited to 550 meters based on the settings of Lucent WaveLAN DSSS radio interface. Two-ray ground radio propagation model is assumed. The channel bit rate is set to 11 Mbps and RTS/CTS access method is used in the simulations. Physical layer preamble and header are transmitted at 1 Mbps, thus, have a total length of 192s according to IEEE 802.11 standard (with Direct Sequence Spread Spectrum) [1].4 Packet payload size is 512 bytes. We use Constant Bit Rate traffic and traffic rate is aggressive enough to keep a contending station always backlogged. The total number of contending stations (N) is increased from 1 to 256. Each simulation lasts for 30 seconds and the presented results are averaged over 20 runs. We measure the results of DSCR and 802.11 in terms of aggregate throughput, average access delay and access energy cost. In the simulated scenario, taking into account the overhead introduced by data packet header (48 bytes), RTS (20 bytes), CTS (14 bytes), ACK (14 bytes), DIFS (50s), SIFS (10s), physical layer preamble and header (192s), respectively, for each of RTS, CTS, DATA and ACK, the total transmission time is 1; 290:18s for each payload packet (512 bytes). Therefore, without any cost of channel contention resolution, we can expect the ideal throughput to be 3,100.3 Kbps. The aggregate throughput of DSCR and 802.11 shown below is normalized to the ideal throughput. The average access delay is measured as the delay experienced by a packet from the time it arrives at MAC layer to the time the source station receives acknowledgment from the destination. To measure the energy consumption, we assume the power consumption model of 2.4 GHz DSSS Lucent IEEE 802.11 WaveLAN PC Card operating in ad-hoc mode with channel bit rate of 11 Mbps 4. In this paper, we do not consider the case that different portions of a packet may have different transmission ranges due to their different transmission rates.
YANG AND H. VAIDYA: A WIRELESS MAC PROTOCOL USING IMPLICIT PIPELINING
TABLE 1 Power Consumption of Lucent IEEE 802.11 WaveLAN PC Card
[21], as shown in Table 1. As both successful and unsuccessful transmission attempts consume energy, a MAC protocol with more retransmissions will be less energy-efficient. We define the “access energy cost” as the total consumed energy by all stations divided by the aggregate throughput (unit: Joule/Kbps).
5.1
Throughput, Access Delay, and Access Energy Cost The aggregate throughput of DSCR and IEEE 802.11 is normalized with respect to the ideal throughput (3,100.3 Kbps) and is presented in Fig. 9a. The standard deviation of the normalized throughput for DSCR and 802.11 is less than 0.0013 and 0.0015, respectively. As we can see, 802.11 has the closest performance to DSCR when there are four contending stations, where 802.11 obtains 91 percent of ideal throughput and DSCR obtains 92 percent. With fewer or greater number of contending stations, the performance gap between DSCR and 802.11 becomes larger. Particularly, with 32 contending stations, DSCR achieves 93 percent of ideal throughput while 802.11 gets 84 percent. When N is 256, DSCR still retains 88 percent of the ideal throughput. Meanwhile, 802.11 has dropped its normalized throughput to 68 percent.5 DSCR not only improves the channel utilization over 802.11 DCF, but as shown in Figs. 9b and 9c, the average access delay and access energy cost of DSCR is also better than 802.11 DCF. For a distributed contention resolution algorithm, the access delay of a packet consists of channel waiting time before its transmission attempts and the time spent in packet retransmissions. In DSCR, many retransmissions are avoided due to reduced collision probability. On the other hand, the channel waiting time does not increase due to the pipelining procedure of DSCR. These two features together explain why DSCR has shorter average access delay while improving the channel utilization for large networks. When N is 256, the average access delay of 802.11 is 0.466 s, while DSCR has average access delay of 0.376 seconds. The standard deviation of the measured access delay for DSCR and 802.11 is 0.001 and 0.019 (seconds), respectively. The reduced collision probability of DSCR also contributes to its reduced access energy cost. As many retransmissions are avoided, each successfully delivered packet costs fewer transmission attempts, thus consumes less energy. When N is 256, the access energy cost of 802.11 DCF is 3.25 Joule/Kbps while that of DSCR is 2.48 Joule/Kbps. The standard deviation of the measured access energy cost for DSCR and 802.11 is 0.003 and 0.077 (Joule/Kbps), respectively. 5. In order to provide meaningful measurements for access delay and access energy cost, in all simulations of this paper, there is no retransmission limit for each packet. On the other hand, the results in [16], which show a larger improvement of DSCR over 802.11 in terms of aggregate throughput, are obtained with short retransmission limit of 7 (as in 802.11).
265
5.2 Key Properties of DSCR To explain why DSCR works better, we take a closer look at the behavior of DSCR with regard to the contention degree in stage 2 and the number of collisions experienced. The total number of contending stations is increased from 1 to 256, and we plot the average number of pipelined stations in stage 2, as shown in Fig. 10a. Given the range of N from 1 to 256, the number of pipelined stations is controlled to be within the range from 1 to 28, which confirms that DSCR controls the channel contention in stage 2 effectively. In the Appendix, which can be found on the Computer Society Digital Library at http://computer. org/tmc/archives.htm, we give theoretical analysis to deduce the average number of pipelined stations in stage 2. Since only pipelined stations in stage 2 contend for the channel at any given time, we expect the collision probability using DSCR can be reduced compared with 802.11. Using RTS/CTS access method with ideal channel conditions, RTS retransmissions are the direct results of collisions. We count the average number of RTS retransmissions per second for DSCR and 802.11, and plot the results in Fig. 10b. As we can see, the number of RTS retransmissions of DSCR is much less than 802.11. Particularly, with 256 contending stations, there is an average of 1461 RTS retransmissions per second using 802.11 and only 414 using DSCR. The results reveal that through the implicitly pipelined contention resolution phase 1, DSCR reduces collision overhead dramatically.
6
PERFORMANCE EVALUATION OF DSCR MULTIHOP WIRELESS NETWORKS
IN
In wireless LANs, all stations can hear each other’s transmission. On the other hand, in multihop networks, two hidden stations that cannot receive from each other may still compete for the channel. Lack of correct channel feedback due to “hidden terminal” problem causes difficulties in efficient contention resolution in multihop networks. Some prior work [22], [23], [24] discuss the problems of 802.11 in multihop networks, mainly dealing with the interactions among physical layer, MAC layer, routing layer and transport layer. The proposed DSCR is not intended to address all these problems existing in multihop networks. Instead, the efforts of DSCR are focused on how to improve the channel utilization via more efficient contention resolution, and various multihop scenarios are evaluated in this section. Recall that the effective transmission range in the simulations is limited to 250 meters and carrier sense range is limited to 550 meters.
6.1 Saturated Multihop Random Networks In this section, 30 different topologies are generated by placing 80 stations randomly in a 1; 000m 1; 000m area. In the simulations, for each generated topology, each station picks one of its one hop neighbors (if there is any) to send packets to. The total number of flows varies from 70 to 75 depending on the corresponding topology.6 Fig. 13 shows 6. In Section 5, we see that the DSCR achieves greater performance improvement in larger networks. To avoid exaggerating the performance gain of DSCR, we choose a moderate network size here.
266
IEEE TRANSACTIONS ON MOBILE COMPUTING,
VOL. 5,
NO. 3,
MARCH 2006
Fig. 9. Throughput, access delay, and access energy cost. (a) Normalized throughput. (b) Average access delay. (c) Access energy cost.
one example of the generated flow patterns. Each source station is always backlogged. Defining “throughput ratio” as the aggregate throughput of DSCR, divided by the aggregate throughput of 802.11, we show the results of “throughput ratio” for the 30 random topologies in Fig. 11a using payload packet size of 512 bytes and the RTS/CTS access method. DSCR achieves 10 percent to 49 percent more throughput compared with 802.11 in these multihop networks (i.e., throughput ratio is in the range from 1.10 to 1.49). The standard deviation of “throughput ratio” is less than 0.047. The average number of RTS retransmissions (per second) for DSCR and 802.11 are plotted in Fig. 11b. We can see that the number of retransmissions experienced by DSCR is much smaller than 802.11, which confirms that DSCR gains more throughput by effectively reducing the collision overhead. Average access delay and energy cost are reported in Figs. 12a and 12b, respectively. DSCR achieves 11 percent to 69 percent less average access delay for each delivered packet, and spends 9 percent to 46 percent less energy per Kbps of obtained throughput. The standard deviation of access delay for DSCR and 802.11 is less than 0.0014 and
0.0008 seconds, respectively. The standard deviation of access energy cost for DSCR and 802.11 is less than 0.0051 and 0.0035 Joule/Kbps, respectively. Comparing with the results for wireless LANs in Fig. 9, we notice that a larger performance improvement over 802.11 can be expected from DSCR in multihop networks. The reason is that, in multihop networks, two source stations that can interfere with each other may not be able to sense each other due to their distance or communication obstacles. Hence, the time period during which a transmission is vulnerable to the possible collisions lasts not only the propagation delay, but also the packet transmission duration (e.g., RTS transmission duration in the case of RTS/CTS access method). As a result, compared with 802.11, the reduced channel contention using DSCR has larger impact on reducing the collision probability, which leads to more performance improvement.
6.2 Multihop Networks with Various Load Having shown the performance improvement of DSCR over 802.11 for saturated networks, we further identify its performance in networks that have less traffic load. The
YANG AND H. VAIDYA: A WIRELESS MAC PROTOCOL USING IMPLICIT PIPELINING
267
Fig. 10. Properties of DSCR. (a) Average number of pipelined stations in stage 2. (b) Average number of RTS retransmissions (per second).
Fig. 11. Throughput and retransmissions in random topologies. (a) Throughput ratio. (b) Number of retransmissions.
purpose of this set of simulations focuses on comparing the performance of DSCR with 802.11 when varying the traffic load from low to high. Poisson traffic with various packet arrival rate is used, and the traffic is generated at the application layer. The length of interface queue between the link layer and the MAC layer is set to one to mitigate the impact of packet buffering at the interface queue. Note that the arrival traffic at the MAC layer is not required to follow any specific distribution since neither DSCR nor 802.11 exploits any particular traffic property to aid the contention resolution. We perform simulations for one of the randomly generated topologies, as shown in Fig. 13. There are a total of 80 stations and 74 one hop flows. Each source station generates its packets independently and the packet arrival rate at each station is (unit: packets per second). With ¼ 1 (i.e., one packet arrives per second for each flow) and 512 byte packet size, the traffic demand is far below the network capacity. When gradually varying from 1 to 20,000 packets/second, offered load is increased from small to very large. The corresponding “aggregate throughput,” “average access delay,” and “access energy cost” are presented in Fig. 14. When the network load is very low, channel contention is not a major concern for the performance, DSCR behaves very similar to 802.11 DCF. Their performance starts to diverge when the network is
loaded more heavily. In this particular example, the aggregate throughput grows slowly beyond ¼ 100 packets/second using 802.11 while DSCR continues to deliver more traffic until ¼ 1; 000 packets/second. The saturation throughput of DSCR is 1.41 times of 802.11. It is interesting to observe that, when ¼ 100 packets/ second, both 802.11 DCF and DSCR have the largest access delay. When offered load increases further, the access delay drops. The main reason is because of the location dependent contention in multihop networks. When ¼ 100, the contending channels are not saturated, but the collision probability among competing flows is high because of hidden terminals (Fig. 13 shows that many hidden terminals exist). Many packets are delivered after multiple retransmissions and, thus, the average access delay is large. When traffic arrival rate increases further, it is likely that some of the flows with less neighborhood contention will grab more channel resource, while the other flows with more severe contention will get fewer packets delivered (i.e., location dependent contention of multihop networks). Since access delay only accounts for the delay of successfully delivered packets, the overall average access delay is decreased. At saturation state, the average access delay of DSCR is 0.025 seconds, while 802.11 DCF has delay of 0.040 seconds.
268
IEEE TRANSACTIONS ON MOBILE COMPUTING,
VOL. 5,
NO. 3,
MARCH 2006
Fig. 12. Access delay and access energy cost in random topologies. (a) Average access delay. (b) Access energy cost.
When network load is low, most of the energy is consumed at idle state, which leads to the measured high energy cost in terms of Joule/Kbps for both DSCR and 802.11. At saturation state, the access energy cost of DSCR is 0.23 Joule/Kbps, while that of 802.11 is 0.32 Joule/Kbps. For above simulation results with ¼ 20; 000 packets/ second, the standard deviation of “aggregate throughput” for DSCR and 802.11 is 80 and 35 Kbps, respectively; the standard deviation of “average access delay” for DSCR and 802.11 is 0.0008 and 0.0004 seconds, respectively; the standard deviation of “access energy cost” for DSCR and 802.11 is 0.0016 and 0.0014 Joule/Kbps, respectively.
6.3 Two Simple Scenarios As seen from the results above, DSCR works well in multihop networks despite the presence of hidden terminals. From the description of DSCR, it should be clear that DSCR benefits when a station can hear the successful transmissions of other nearby stations. However, with hidden terminals, such detection may not always be feasible. We now consider two simple scenarios, shown in Fig. 15, to explore the impact of hidden terminals further.
Fig. 13. One random multihop network.
The location of each station is shown in the form of (x, y) coordinates. The first scenario in Fig. 15a has six constantly backlogged source stations 1, 2, 3, 4, 5, and 7, where station 7 is out of the transmission range of stations 1-5, but within the transmission range of station 6. Via CTS or ACK sent by station 6, station 7 can learn of the transmissions from stations 1-5. Using DSCR, when in stage 1, station 7 can reduce its bc1 whenever a transmission from stations 1-5 finishes. Once becoming a pipelined station and contending the channel in stage 2, station 7 can tell whether it has lost the channel or not via the overheard CTS from station 6. As a result, from all contending stations 1, 2, 3, 4, 5, and 7, DSCR only select a subset of them at any point of time to contend for the channel in stage 2. The throughput obtained by each individual flow, the aggregate throughput and the number of RTS retransmissions occurred per second are presented in Fig. 15a for both DSCR and 802.11 DCF. Notice that, when station 7 collides with one of the stations from 1 to 5, the packet reeception at station 6 will be interfered, but the packet reception at station 8 can proceed successfully, which explains why flow 6 has much larger throughput than other flows using 802.11 DCF. On the other hand, when DSCR is used, station 7 does not always get chances to contend for the channel access in stage 2, which mitigates its channel access advantage over others to some extent. The second scenario in Fig. 15b is almost identical to the first scenario except that station 7 is further moved out of the transmission range of station 6. Now, station 7 cannot overhear any successful transmissions from stations 1-6, but it is still within their carrier sense range. In DSCR, a station has two ways of reducing bc1 . One way is through overheard successful transmissions, the other is to reduce bc1 by 1 after each channel idle slot. Therefore, even though station 7 cannot overhear any successful transmissions, it can still enter stage 2 and contend for the channel when its bc1 reaches zero via the second way. At the same time, DSCR only selects a subset from stations 1-5 to become pipelined stations and enter stage 2 at any point of time. When contending for the channel in stage 2 with other
YANG AND H. VAIDYA: A WIRELESS MAC PROTOCOL USING IMPLICIT PIPELINING
269
Fig. 14. Poisson traffic in random topologies (horizontal axis uses log-scale). (a) Aggregate throughput. (b) Average access delay. (c) Access energy cost (Y axis uses log-scale).
pipelined stations, station 7 does not know whether it has lost the channel access in stage 2 or not without overhearing any successful transmissions. Consequently, station 7 will not return to stage 1 until it has won the channel access and transmitted a packet, which explains why flow 6 has larger throughput than others in Fig. 15b using DSCR. On the other hand, when using 802.11, station 7 contend for the channel with all the stations 1-5. Whenever the transmissions from stations 1-5 are sensed, station 7 has to defer its transmissions, which lead to a relatively small throughput of flow 6 compared with other competing flows. The key observation from both examples is that, despite the presence of hidden terminals in multihop networks, DSCR is still able to select only a subset of contending stations to become pipelined stations and contend for the channel access in stage 2. Due to the reduced channel contention, DSCR experiences less number of RTS retransmissions and has better throughput than 802.11, even though the difference is not significant due to the limited number of flows in the chosen scenarios. When the number
of contending flows is increased, the throughput benefit of DSCR over 802.11 becomes more significant.
7
RELATED WORK AND WITH DSCR
SOME COMPARISONS
Contention resolution for multiple access control has been extensively studied for wired networks and infrastructure based wireless networks (i.e., networks in which mobile stations communicate via the base station). One category of the proposed algorithms (e.g., [11], [12]) makes use of ternary channel feedback (i.e., channel idle (0), collision (e), and successful transmission (1)) to estimate the network load and stabilize the contention resolution algorithm. Some other tree splitting protocols (e.g., [13], [25], [26]) try to resolve collisions fast and improve the throughput. In tree splitting protocols, when a collision happens, the colliding stations are split into two subsets based on a certain rules, whereas the first subset transmits in the next slot. The stations in the second subset must wait until all the stations in the first subset have succeeded. If the collision is not
270
IEEE TRANSACTIONS ON MOBILE COMPUTING,
VOL. 5,
NO. 3,
MARCH 2006
Fig. 15. Example scenarios for multihop networks. (a) Scenario 1. (b) Scenario 2.
resolved, then a further splitting into subsets takes place. Unfortunately, in tree splitting algorithms, a station also needs ternary channel feedback to get hold of the contention resolution progress. In wireless networks, especially multihop networks, channel feedback is not reliable in terms of reflecting actual medium status due to the well-known “hidden terminal” and “exposed terminal” problems. Consequently, the existing tree splitting protocols and stabilization techniques cannot be used [14], and researchers continue to search for efficient contention resolution algorithms for wireless networks. Among them, some prior research work, e.g., [27], proposes to use one common subchannel to schedule packets for multiple data subchannels. Compared with the pipelined packet scheduling schemes discussed in this paper, the fundamental difference lies in that, in the prior schemes, the exchange of control messages to decide which station will transmit on an available data channel occurs when at least one of the channels is perceived as idle. Contrary to this, the contention resolution of the pipelined schemes proceeds for packets to be transmitted in the future when the channel is currently busy. Some reservation schemes are also proposed (e.g., [28]). To use a reservation scheme, a station has to know beforehand when the next packet will arrive, which may not be feasible in many applications. Additionally, contending stations within a station’s two-hop neighborhood need to be aware of the channel reservation made by this station, which involves significant overhead when the network topology changes often due to mobility. Yuang et al. [29] proposes a hexanary-feedback contention access scheme for infrastructure based wireless networks. In their proposed scheme, mobile stations are allowed to have full knowledge of the exact number of
users involved in a collision with the support of hardware. A two-phase contention resolution process is then used to efficiently resolve channel contention. Tay et al. [30] notice that some wireless networks have the bursty event-based traffic patterns and propose a nonpersistent carrier sense multiple access protocol that attempts to minimize collisions between contending stations for such networks. Cali et al. [9], [10] propose a stabilized “Dynamic 802.11” algorithm specifically for a wireless LAN, which is a special type of wireless networks in which all stations can hear each other’s transmission and contend for a common channel. “Dynamic 802.11” estimates the channel idle time and collision time from the observed channel activity, thus, minimizes the contention resolution overhead. However, due to the location-dependent contention and hidden terminal problems in multihop networks, local channel activities observed do not reflect the actual status of the contended channel. Additionally, “Dynamic 802.11” only deals with the situations where all contending stations always have packets ready for transmissions. [31] proposes an enhanced p-persistent IEEE 802.11 protocol to achieve both the channel access efficiency and energy saving. Other contention resolution optimization schemes for 802.11 DCF include [32], [33]. The advantages of DSCR when compared with existing work lie in that, by applying pipelining techniques, DSCR is able to adapt to the limited available channel feedback in wireless networks and improves the contention resolution efficiency over 802.11 DCF in both wireless LANs and multihop networks. In the rest of this section, we discuss “Dynamic 802.11” [9], [10] and a fast collision resolution algorithm (FCR) [33] in more details.
YANG AND H. VAIDYA: A WIRELESS MAC PROTOCOL USING IMPLICIT PIPELINING
271
Fig. 16. Comparison of dynamic 802.11 with DSCR and IEEE 802.11 (without RTS/CTS, payload size: 512 bytes).
7.1 Discussion of “Dynamic 802.11” Some prior research work (e.g., [5], [7], [9], [10]) proposes to tune the contention window of 802.11 to improve its performance in wireless LANs. Particularly, the “Dynamic 802.11” proposed in [9] dynamically adjusts each station’s transmission probability based on estimated channel idle time and collision time. In [9], it is argued that the network can operate at a point close to the optimal if the channel time wasted on idle periods is equal to the channel time spent on collisions. Based on this argument, [9] proposes a scheme in which each station continuously observes the channel activities to estimate the average idle/collision period (“Dynamic 802.11” is proposed for basic access method of 802.11 and a channel busy period is assumed to be a collision if an ACK does not immediately follow). Then at the end of every transmission attempt, each station computes its current estimation for the number of contending stations and the optimal value of transmission probability using the derived analytical results in [9]. The optimal transmission probability can be mapped to the average contention window value for each station, and the channel utilization of “Dynamic 802.11” can be optimized. Intuitively, in “Dynamic 802.11,” a station will increase its transmission probability when observed idle period is longer than collision period and decrease its transmission probability when the idle period is shorter than the collision period. However, the optimality of “Dynamic 802.11” relies on stations’ capability of obtaining channel status correctly. “Missed ACK,” “Carrier Sensing Fault,” and “Not-detected Transmissions” will cause idle periods/collision periods to be overestimated or underestimated, as noted in [10] as well. Especially in multihop networks, incorrect observation of channel status will fundamentally mislead the algorithm and result in poor performance of “Dynamic 802.11.” The difference between “Dynamic 802.11,” DSCR, and 802.11 is further illustrated in Fig. 16 using simulation results for three simple scenarios. We use the basic access method (i.e., without the RTS/ CTS handshake) because currently “Dynamic 802.11” is designed only for the basic access method [9]. Recall that transmission range is assumed to be 250 meters, and carrier
sensing range is 550 meters for all four stations in the simulations. The first one is a typical wireless LAN scenario, in which all four stations can hear each other’s transmissions correctly. In this case, “Dynamic 802.11” reaches its optimal point and achieves the best channel utilization, followed by DSCR and then 802.11. In the second scenario, four stations are still within each other’s carrier sensing range, but station 1 cannot correctly receive the packets transmitted by stations 3 and 4. Similarly, station 4 cannot correctly receive the packets transmitted by stations 1 and 2. As shown in Fig. 16, the result for “Dynamic 802.11” is that one of the two source stations (station 4 in this particular simulation run) will grab the channel completely and the other one will be starved. The reason is explained as follows: During a successful packet transmission from station 4, station 1 can sense the DATA transmission from station 4 and ACK transmission from station 3, but can correctly receive neither of them. Therefore, station 1 considers the transmission from station 4 as collision, and its transmission probability will be decreased accordingly. On the other hand, station 4 can receive ACK from station 3 correctly and then increases its transmission probability. Consequently, station 4 has more chances than station 1 to transmit next packet. Next time when station 4 transmits, station 4 will again increase its transmission probability and station 1 will again decrease its transmission probability. In this way, station 4 eventually grabs the channel and station 1 is starved due to over-estimation of collision time. DSCR and IEEE 802.11 do not have such a problem, and DSCR achieves more throughput than 802.11, as the simulation results show. In the third scenario in Fig. 16, stations 1 and 4 are moved further away and they cannot sense each other’s transmissions anymore. When station 4 is transmitting, station 1 will sense the channel as idle. Since station 1 overestimates the channel idle period, it tends to transmit more aggressively than it should. A similar case happens to station 4 as well. When stations 1 and 4’s transmissions overlap in time, enough interference is caused at stations 2 and 3 and collisions occur. The collision probability is high
272
IEEE TRANSACTIONS ON MOBILE COMPUTING,
since both stations 1 and 4 have inappropriately high transmission probability on average. The performance of “Dynamic 802.11” degrades dramatically compared with its performance in scenario 1. The primary problem of “Dynamic 802.11” is that it entirely relies on the estimated channel idle time and collision time to calculate the transmission probability. If complete and accurate channel feedback is available, it can achieve the optimal performance. Otherwise, its performance can be quite poor being misled by incorrect channel feedback. On the other hand, using DSCR or IEEE 802.11, stations 1 and 4 mainly exploit channel feedback from their own transmissions (success or failure) to adjust the contention window. When collisions occur, stations 1 and 4 will double their contention window, thus reduce the collision probability for next transmission attempt. The probabilistic procedure can resolve the contentions between stations 1 and 4 most of the time, even though the collision probability is higher than scenarios 1 and 2 due to the lack of ability for stations 1 and 4 to sense each other’s data transmissions. As the simulation results show, the performance degradation of DSCR and IEEE 802.11 is much less than “Dynamic 802.11” in this scenario. In this case, DSCR achieves more throughput than both 802.11 and “Dynamic 802.11.” In summary, “Dynamic 802.11” does not perform well in multihop networks, compared with the proposed implicitly pipelined DSCR protocol.
7.2 Discussion of Fast Collision Resolution (FCR) Kwon et al. [33] have proposed a Fast Collision Resolution (FCR) algorithm for wireless LANs. With FCR, whenever a new busy period is detected (could be either a collision or a packet transmission), all deferring stations will exponentially increase their own contention window and generate a new backoff counter. In this way, all deferring stations have larger contention window compared with the winning station; thus, the winning station has higher probability to win channel access again. As FCR reduces the collision probability by letting one of the contending stations occupy the channel for a certain period of time, unfairness issue arises. Comparing with 802.11, fairness among contending stations could get worse and worse as the number of contending stations increases. To overcome this issue, FCR defines a maximum transmission limit and uses distributed SCFQ algorithm to dynamically adjust this maximum transmission limit. If a station consecutively succeeds in channel access over this transmission limit, it will change its contention window to a maximum value in order to give other stations opportunities to transmit. The fundamental difference between FCR and DSCR is that, FCR reduces collision probability by letting one station occupy the channel for a relatively long time, while DSCR reduces collision probability by statistically selecting a subset of stations to contend for the channel at any given time. As the group of pipelined stations staying in stage 2 of DSCR generally change after every successful transmission, DSCR is able to maintain comparable fairness to 802.11. Currently, FCR is proposed only for wireless LANs [33]. As FCR uses SCFQ fair scheduling algorithm to resolve the unfairness issue, it is not clear how to extend FCR further to multihop networks. On the other hand, DSCR can work
VOL. 5,
NO. 3,
MARCH 2006
well both in wireless LANs and multihop networks, as we showed earlier using simulation results.
8
CONCLUSION
Pipelined packet scheduling is discussed in this paper. In particular, a MAC protocol—DSCR (pipelined Dual-Stage Contention Resolution)—that uses an implicitly pipelined contention resolution algorithm is proposed. By applying implicit pipelining in DSCR, most of the channel idle overhead is hidden and the channel contention is reduced without introducing any additional cost. Extensive simulation results show that DSCR significantly improves the performance of 802.11 DCF in both wireless LANs and multihop networks in terms of channel utilization, average access delay and access energy cost. Overall, without dependence on extensive channel feedback information (e.g., as in “Dynamic 802.11” [9]), without relying on elimination burst or other signaling mechanisms (e.g., as in HIPERLAN/I [19]), DSCR provides an improved performance for IEEE 802.11 DCF and, hence, demonstrates that pipelining techniques can be useful in improving the performance of multiple access control protocols.
ACKNOWLEDGMENTS The authors would like to thank the referees for their helpful comments. Preliminary versions of this paper appeared at the ACM International Symposium on Mobile Ad Hoc Networking and Computing (Mobihoc ’03) poster session and IEEE Semiannual Vehicular Technology Conference Fall 2003 (invited paper). This research is supported in part by US National Science Foundation grant 01-96410.
REFERENCES [1] [2] [3] [4] [5]
[6]
[7] [8]
[9]
“Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications,” IEEE Standard 802.11, June 1999. H.S. Chhaya and S. Gupta, “Performance Modeling of Asynchronous Data Transfer Methods of IEEE 802.11 MAC Protocol,” Wireless Networks, vol. 3, pp. 217-234, 1997. B.P. Crow, I. Widjaja, J.G. Kim, and P.T. Sakai, “IEEE 802.11 Wireless Local Area Networks,” IEEE Comm. Magazine, vol. 2, pp. 116-126, Sept. 1997. Y.C. Tay and K.C. Chuan, “A Capacity Analysis for the IEEE 802.11 MAC Protocol,” ACM/Baltzer Wireless Networks, vol. 7, no. 2, pp. 159-171, Mar. 2001. J. Weinmiller, H. Woesner, J.P. Ebert, and A. Wolisz, “Analyzing and Tuning the Distributed Coordination Function in the IEEE 802.11 DFWMAC Draft Standard,” Proc. Int’l Workshop Modelling MASCOT, Feb. 1996. J. Weinmiller, M. Schlager, A. Festag, and A. Wolisz, “Performance Study of Access Control in Wireless LANs IEEE 802.11 DFWMAC and ETSI RES 10 HIPERLAN,” Mobile Networks and Applications, vol. 2, pp. 55-67, 1997. G. Bianchi, “Performance Analysis of the IEEE 802.11 Distributed Coordinated Function,” IEEE J. Selected Areas in Comm., vol. 18, no. 3, pp. 535-547, Mar. 2000. G. Bianchi, L. Fratta, and M. Oliveri, “Performance Evaluation and Enhancement of the CSMA/CA MAC Protocol for 802.11 Wireless LANs,” Proc. Ann. IEEE Int’l Symp. Personal Indoor and Mobile Radio Comm. (PIMRC), vol. 2, pp. 392-396, Oct. 1996. F. Cali, M. Conti, and E. Gregori, “IEEE 802.11 Protocol: Design and Performance Evaluation of an Adaptive Backoff Mechanism,” IEEE J. Selected Areas in Comm., vol. 18, no. 9, pp. 1774-1786, Sept. 2000.
YANG AND H. VAIDYA: A WIRELESS MAC PROTOCOL USING IMPLICIT PIPELINING
[10] F. Cali, M. Conti, and E. Gregori, “Dynamic Tunning of the IEEE 802.11 Protocol to Achieve a Theoretical Throughput Limit,” IEEE/ ACM Trans. Networking, vol. 8, no. 6, pp. 785-799, Dec. 2000. [11] B. Hajek and T. VanLoon, “Decentralized Dynamic Control of a Multiaccess Broadcast Channel,” IEEE Trans. Automatic Control, 1982. [12] R.L. Rivest, “Network Control by Bayessian Broadcast,” Technical Report MIT/LCS/TM-285, Laboratory for Computer Science, Massachusetts Inst. of Technology, Cambridge, Mass., 1985. [13] J.I. Capetanakis, “Tree Algorithms for Packet Broadcast Channels,” IEEE Trans. Information Theory, vol. 25, no. 5, pp. 505-515, Sept. 1979. [14] D. Bertsekas and R. Gallager, Data Networks, second ed. Prentice Hall, 1992. [15] J. Hennessy and D. Patterson, Computer Architecture: A Quantitative Approach, third ed. Morgan Kaufmann, 2002. [16] X. Yang and N.H. Vaidya, “Explicit and Implicit Pipelining for Wireless Medium Access Control,” Proc IEEE Semiann. Vehicular Technology Conf., Fall 2003. [17] X. Yang and N.H. Vaidya, “Pipelined Packet Scheduling in Wireless LANs,” technical report, Coordinated Science Laboratory, Univ. of Illinois at Urbana-Champaign, Aug. 2002. [18] J. Deng, Y.S. Han, and Z.J. Haas, “Analyzing Split Channel Medium Access Control Schemes with ALOHA Reservation,” Proc. Second Int’l Conf. AD-HOC Networks and Wireless (ADHOCNOW ’03), Oct. 2003. [19] ETSI TC-RES, “Radio Equipment and Systems (RES); HIgh PErformance Radio Local Area Network (HIPERLAN) Type 1; Functional Specification,” Oct. 1996, European Telecomm. Standard ETS 300 652. [20] VINT Group, “UCB/LBNL/VINT Network Simulator NS (Version 2),” http://www.isi.edu/nsnam/ns, 1996. [21] L.M. Feeney and M. Nilsson, “Investigating the Energy Consumption of a Wireless Network Interface in an Ad Hoc Environment,” Proc. IEEE INFOCOM Conf., 2001. [22] Z. Fu, P. Zerfos, H. Luo, S. Lu, L. Zhang, and M. Gerla, “The Impact of Multihop Wireless Channel on TCP Throughput and Loss,” Proc. IEEE INFOCOM Conf., 2003. [23] S. Xu and T. Saadawi, “Does the IEEE 802.11 Mac Protocol Work Well in Multihop Wireless Ad Hoc Networks,” IEEE Comm. Magazine, vol. 39, no. 6, pp. 130-137, June 2001. [24] S. Xu and T. Saadawi, “Revealing the Problems with 802.11 Medium Access Control Protocol in Multihop Wireless Ad Hoc Networks,” Computer Networks, vol. 38, no. 4, pp. 531-548, Mar. 2002. [25] R.G. Gallager, “Conflict Resolution in Random Access Broadcast Networks,” Proc. AFOSR Workshop Comm. Theory and Applications, Sept. 1978. [26] B.W. Wah and X. Su, “An Efficient Multiaccess Protocol for Wireless Networks,” Proc. Int’l Symp. Internet Technology, pp. 173178, Apr. 1998. [27] A. Tzamaloukas and J.J. Garcia-Luna-Aceves, “A ReceiverInitiated Collision-Avoidance Protocol for MultiChannel Networks,” Proc. IEEE Infocom Conf., 2001. [28] G.R. Hiertz, J. Habetha, P. May, E. Weiss, R. Bagul, and S. Mangold, “A Decentralized Reservation Scheme for IEEE 802.11 Ad Hoc Networks,” Proc. IEEE Personal Indoor Mobile Radio Conf. (PIMRC), Sept. 2003. [29] M.C. Yuang, B.C. Lo, and J.-Y. Chen, “Hexanary-Feedback Contention Access with PDF-Based Multiuser Estimation for Wireless Access Networks,” IEEE Trans. Wireless Comm., vol. 3, no. 1, pp. 278-289, Jan. 2004. [30] Y.C. Tay, K. Jamieson, and H. Balakrishnan, “Collision-Minimizing CSMA and Its Applications to Wireless Sensor Networks,” IEEE J. Selected Areas in Comm., vol. 22, no. 6, pp. 1048-1057, Aug. 2004. [31] R. Bruno, M. Conti, and E. Gregori, “Optimization of Efficiency and Energy Consumption in p-Persistent CSMA-Based Wireless LANs,” IEEE Trans. Mobile Computing, vol. 1, no. 1, pp. 10-31, Jan. 2002. [32] H. Kim and J.C. Hou, “Improving Protocol Capacity with ModelBased Frame Scheduling in IEEE 802.11-Operated Wlans,” Proc. ACM MobiCom, Sept. 2003. [33] Y. Kwon, Y. Fang, and H. Latchman, “A Novel Medium Access Control Protocol with Fast Collision Resolution for Wireless LANs,” Proc. Infocom Conf., 2003.
273
Xue Yang received the BE and MS degrees from University of Electronic Science and Technology of China and the PhD degree from the University of Illinois at Urbana-Champaign (UIUC). She is awarded Vodafone-U.S. Foundation Graduate Fellowship from 2003 to 2005. Her current research is in the areas of wireless networking and mobile computing, with the focus on medium access control, quality of service, and topology control. For more information, please visit http://www.crhc.uiuc.edu/~xueyang/. She is a member of the IEEE. Nitin H. Vaidya received the PhD degree from the University of Massachusetts at Amherst. He is presently an associate professor of electrical and computer engineering at the University of Illinois at Urbana-Champaign (UIUC). He has held visiting positions at Microsoft Research, Sun Microsystems, and the Indian Institute of Technology-Bombay. His current research is in the areas of wireless networking and mobile computing. His research has been funded by various agencies, including the National Science Foundation, DARPA, BBN Technologies, Microsoft Research, and Sun Microsystems. He is a recipient of a CAREER award from the US National Science Foundation. He has served on the program committees of several conferences and workshops, and served as program cochair for the 2003 ACM MobiCom. He has served as editor for several journals, and presently serves as editor-in-chief of the IEEE Transactions on Mobile Computing. For more information, please visit http://www.crhc.uiuc.edu/~nhv/. He is a senior member of the IEEE and the IEEE Computer Society.
. For more information on this or any other computing topic, please visit our Digital Library at www.computer.org/publications/dlib.
