SPLASH: a Simple Multi-Channel Migration Scheme for IEEE Networks

Transcription

1 SPLASH: a Simple Multi-Channel Migration Scheme for IEEE Networks Seungnam Yang, Kyungsoo Lee, Hyundoc Seo and Hyogon Kim Korea University Abstract Simultaneously utilizing multiple channels can be an effective counter-measure to the performance degradation stemming from high contention in the IEEE based networks. This is especially true in multi-hop networks where spatial interference exacerbates the contention problem. Existing proposals use a special channel, a time window, or an orchestrated channel switching schedule to coordinate the spreading of the stations across channels. In this paper, we propose a simpler scheme called Splash, which is mostly built on the current standard. The core idea of the Splash scheme is simply that whenever a pair of stations starts communication on a channel, all other stations jump to another channel next in priority. Simulation and analysis show that this simple scheme boosts the throughput by 4 to 5 times, with a larger reduction in delay. I. INTRODUCTION It is well known that the IEEE MAC are subject to high collision probability. For instance, the single-hop 82.11b wireless LAN (WLAN) with 5 contending stations has close to 2% chance of collision for each transmission, which rises to close to 3% in the 82.11a WLAN (Fig. 1) [1]. This problem is exacerbated when the technology is used in multihop wireless networks such as mesh networks. Additional spatial interference from adjacent links worsens the situation, causing significant overall throughput decrease. probability single collision, 82.11a single collision, 82.11b 2 consecutive collisions, 82.11a number of stations Fig. 1. Collision probability vs. number of contending stations in the IEEE 82.11a/b WLAN. One attractive solution approach to the problem is dynamically utilizing the multiple channels provisioned in the standards. For instance, the 82.11a standard specifies 12 orthogonal channels that can be possibly simultaneously used by different members of the WLAN or multi-hop wireless networks. There have been three very different approaches to multiple channel utilization in the IEEE context. Wu et al. propose Dynamic Channel Assignment (DCA) scheme [2] that uses a separate channel for multi-channel access control. Each station is equipped with two half-duplex transceivers, one of which operates in the control channel and the other, in one of other data channels. The shortcomings of this approach are that we need two transceivers, and one channel is exclusively used for signaling. But channel negotiation can be done anytime. So and Vaidya propose the MMAC protocol [3]. Instead of a dedicated control channel, it uses a (default) data channel that takes the role of the control channel. The stations coordinate the channel usage in the Announcement Traffic Indication Message (ATIM) window, but the rest of the time the default channel can still be used for data exchange. So a higher channel utilization is achieved. As the channel negotiation is done in the ATIM window, however, the length of the window can affect the performance. Bahi et al. proposes Slotted Seeded Channel Hopping (SSCH) scheme [4], where each station is assigned a seed that determines the hopping sequence. The sequence is designed to guarantee the stations to meet in a channel periodically. If communication is needed, they meet this way and lock the seed until they finish the transmission. The shortcoming of this approach is that there is a delay until a station meets the intended peer. In this paper, we propose a novel channel migration scheme called Splash that uses neither a special signaling channel nor a special time window nor a carefully orchestrated channel switching schedule to coordinate the channel usage by multiple stations. The core idea of the Splash scheme is simply that whenever a pair of stations starts communication on a channel, all other stations jump to another channel next in priority. The rational is that once a pair of stations begins exchanging frame(s), the other stations have no hope of using that particular channel in the following moments. The migration of the non-communicating stations allows them to initiate communication in the new channel. Migrations can occur iteratively across multiple channels as traffic demand rises, before all stations return to the default channel again. One assumption of the Splash scheme is that the channel priority is predetermined and all stations know it. For instance, in 82.11b, we could have the priority set to , where the channel 6 serves as the default channel. (We do not impose any special, active role to the default channel other than serving as the rallying point for the stations.) Another assumption is that once the communication begins between a pair of stations, the channel occupation is typically longer /8/$ IEEE 2355

2 than a single frame exchange. We believe that in order to avoid suboptimal throughput performance due to framing and channel access overhead in new, higher-speed standards, stations are more likely to make use of the features like frame aggregation (as in the IEEE 82.11n [5]) and Transmission Opportunity (TXOP) (as in the IEEE 82.11e [6]), both of which would mean the channel occupation typically longer than a single frame exchange time. Below, we discuss the design details of the Splash scheme, followed by the pseudo-code and the performance evaluation. II. SPLASH ALGORITHM The TXOP is a newly introduced concept in the IEEE 82.11e standard [6]. It can be used to guarantee or limit the time during which the station occupies the channel. A station can obtain the TXOP through Enhanced Distributed Channel Access (EDCA) or HCF Controlled Channel Access (HCCA). It is made possible by winning the EDCA contention, or by receiving the QoS CF-Poll from the AP, respectively. In an IBSS, which is our multi-hop network, the station that wins the contention gets the TXOP through the EDCA rule [6]. In this paper, we assume that the time duration that a station will actually occupy the channel within the TXOP is conveyed in the Duration/ID field of the first frame. This is quite a stretch, since usually a frame reserves the channel with NAV for only the next frame. We require this change in the use of Duration/ID field so that every other station can set the NAV, or in Splash terms, know by when they should come back to the default channel. Fig. 2. Splash operation example. Fig. 2 shows a simple example of the Splash operation. In this example IBSS are 6 stations in the mutual reception range, on the default channel C 1. Initially, A initiates transmission to B, successfully obtaining the TXOP. All other stations including B pick up the transmission and finds that the transmission is A B. Until the stations figure out that they are (or are not) involved in this transmission, and the actual transmission Duration from the first frame, they remain on C 1, which is marked by WAIT-s in the figure. The waiting time is assumed to be DATA+SIFS+ACK. At this point, as C, D, E, F have no hope of using C 1 during the transmission, they all jump to channel C 2,nextinpriority.OnC 2, both C and E attempt to transmit to D. The transmissions collide, and after backoff, C obtains the channel (TXOP). Here, F detects the activity on the channel, and attempts to figure out that the transmission is successful, but after WAIT-c it knows that it was a collision. The waiting period is assumed to have a length of DATA+DIFS. Then again, E and F migrate to the next channel, C 3, where they initiate communication. There are implementation choices as to when the other stations change channels. An early jump is possible when they parse the header and figure out they are not the intended receiver. A late jump is what is shown in the above example. In this paper, we take the latter approach, since in case of collision, the intended receiver may not realize that it is one of the intended receivers of the collided transmissions (because in that case, the station should not jump to other channel). We will investigate the early jump because it can enable the Splash scheme to be effective even when single frame exchanges dominate. Algorithm 1-3 show the pseudo-code of the Splash algorithm. In this paper, we simply assume that the TXOP on C 1 is fully utilized 1. And at the end of this TXOP, all stations including those that have migrated must converge on C 1 again. Then the whole Splash process as illustrated in Fig. 2 starts again. So the TXOP in C 1 is the master timer that every station keeps and refers to. Now we discuss the algorithms starting from Algorithm 1. In case the remaining time at C 1 is estimated to be less than a single frame exchange duration, no station can initiate a transmission in any channel but should get ready to return to C 1. This inhibition is marked by the state BLOCKED. Otherwise, the backlogged stations attempt to transmit if the CSMA/CA backoff has been performed, through SPLASH TRANSMITTER. And other stations run SPLASH OTHERS. Once T R hits, it is that time to go back to the default channel, so jump is made to C 1 and the state is reset to READY. Upon starting Algorithm 2, the SPLASH TRANSMITTER, the station transmits the first frame and checks if it has succeeded. If not, the whole frame time to clear the failed frame from the channel has been wasted. In this paper, we assume that the time is DATA+DIFS. Then the binary exponential backoff (BEB) is performed. In case the initial frame is a success, on the other hand, the state is updated to TRANSMITTING. Since the TXOP value is announced at the beginning of the successful frame through the Duration/ID field, by this time all other stations know how long C 1 will be occupied. When a station receives a transmission, Algorithm 3, SPLASH OTHERS begins to execute. If the transmission is a success and it happens in C 1, T R is set. If the station is the intended receiver of the transmission, the state is changed to RECEIVING. Otherwise, there is no hope of using the current channel until the next TXOP at C 1, so the station should switch the channel. In doing so, a check is made if all channels have been exhausted. Otherwise, the channel next in priority is selected to jump to. Note that this channel switching is executed by all stations neither transmitting nor receiving currently, thereby implementing the essence of the Splash scheme. If the transmission turns out to be a failure, a wait interval is enforced before any decision is made. 1 This assumption is for convenience of illustration. As long as the Duration field can be used to inform other stations of the expected actual channel occupation time, the TXOP can be utilized in any fraction. 2356

3 Algorithm 1 SPLASH(N) /* N : number of usable channels */ /* C : current channel */ /* T R : remaining duration in the current TXOP at C 1 */ /* T s : frame exchange time =DATA+SIFS+ACK+SIFS */ 1: while (1) do 2: if T R <T s then 3: /* Too little time until rallying back at C 1 */ state BLOCKED 4: else 5: if state = ATTMPTING then 6: if backoff = then 7: SPLASH TRANSMITTER 8: else 9: Decrement backoff timer 1: end if 11: else if state = READY then 12: SPLASH OTHERS 13: end if 14: end if 15: Decrement T R 16: if T R =then 17: C 1 18: state READY 19: end if 2: end while Algorithm 2 SPLASH TRANSMITTER() 1: Transmit /* and set TXOP as NAV */ 2: if success then 3: state TRANSMITTING 4: else 5: /* Collision or channel loss */ 6: wait(data + DIFS) 7: Perform BEB 8: end if 9: return The number of possible transmissions in a channel k is a function of the number of channel migrations and the collisions experience in the channel, i.e., k = max(t R /T s, ), T R TXOP. Table I shows the maximum possible value of TXOP/T s at each channel given the number of collisions that have occurred in the channel. Here TXOP is assumed to be 3ms and data frame has the length of 1K bytes. Due to the constraint that all nodes return to C 1 before the TXOP at C 1 ends, k dwindles as the priority of channel decreases or the number of collisions c increases. Note that in C 1, the TXOP begins only when the first communicating pair grabs the channel, T R /T s always the maximum value of (= TXOP/T s ). III. PERFORMANCE EVALUATION A. Analysis Suppose there are n backlogged stations in the mutual reception range. In [7], the throughput S of the DCF Algorithm 3 SPLASH OTHERS(N,C) 1: if transmission then 2: if success then 3: if C=C 1 then 4: T R TXOP T s 5: end if 6: if receiver then 7: /* It s for me */ 8: state RECEIV ING 9: else 1: /* Should switch channel */ 11: if C<Nthen 12: C next(c) 13: else 14: /* No channel to jump to */ 15: state BLOCKED 16: end if 17: end if 18: else 19: /* Transmission failure */ 2: wait(data + DIFS) 21: end if 22: end if 23: return TABLE I T R /T s Channel max (c =) c =1 c =2 c =3 c =4 c =5 C C C C C C C C system with a single-channel operation is given by: S single = P s P tr E[P ] (1 P tr )σ + P tr P s T s + P tr (1 P s )T c, where H denotes the length of the PHY/MAC header, E[P ] the average frame size, δ the small propagation delay, and σ the slot time. P tr =1 (1 τ) n is the probability that there is at least one transmission in a slot given τ is the transmission attempt probability of a station in the slot, and P s is the conditional probability that the transmission succeeds. Namely, nτ(1 τ)n 1 P s = P tr And T c is the time wasted by collision, and is given by T c = H + E[P ]+DIFS + δ. However, when a burst of frames is sent in a TXOP, the transmission time of a frame, T s, is different from [7]. After each frame in the TXOP, a SIFS is given instead of DIFS. So, T s = H + E[P ]+SIFS + δ + ACK + SIFS. 2357

4 In a single-channel scenario, the throughput formula under TXOP is given as follows: S txop = P s P tr (k E[P ]) (1 P tr )σ + P tr P s (k T s )+P tr (1 P s )T c, (1) where k is the number of data frames that we can send in a TXOP. Namely, k = TXOP T s. Now, suppose we have a multi-channel system with N channels. If we extend Eq. (1) for the multi-channel operation of Splash, we get N 1 P s P tr i= S multi = (k i E[P ]), (1 P tr )σ + P tr P s (k T s )+P tr (1 P s )T c where k i now becomes, k i = TXOP i (T h +(1 P tr)σ + P tr(1 P s)t c + P trp sp mct c) T s. Here, T h = T s + T sw is the time it takes a station to hop to another channel. It is composed of the time to confirm the successful transmission of a frame at the start of the TXOP (T s ) and the time to tune the transceiver to the next priority channel (T sw ). Within a TXOP, we exclude the channel hopping overhead (with probability 1), collision overhead (with probability P tr (1 P s ), and the wasted time by attempting to transmit to an intended station in other channel (with probability P tr P s P mc to determine the number of data frames we can send. And the overhead from the backoffs before transmission, (1 P tr )σ is also subtracted. The remaining time determines k i. P mc is the probability of multi-channel collision, namely, the probability of transmission failure due to the fact that the intended receiver is in other, higherpriority channel, communicating. This probability depends on the number of migrations performed so far. It is given by P mc = 2h N 2h, where h is the number of system-wide channel migrations of stations. So the denominator denotes the number of idle stations that can still create a data exchanging pair, whereas the numerator the number of communicating pairs. The intuition is that every channel migration leaves a pair of communicating stations behind, to which no transmission is possible. This condition is the multi-channel collision. Finally, we need p and τ to compute P s and P tr. The collision probability on a single channel without TXOP is given by p =1 (1 τ) n 1. In other words, it is: p =1 {(1 τ) n +(1 τ) n 1 τ} However, in multi-channel, even a single transmission can collide due to multi-channel collision, so we modify the equation to p =1 {(1 τ) n 2h +(1 τ) (n 2h) 1 τ(1 P mc )}. As τ is a function of p, we can numerically solve the equation for both of them [7]. Below, we compare the throughput from this model with the simulation. B. Simulation We assume 82.11a where each link operates at 54Mbps. The channel switching time T sw is 8 µs [8], [9], and the packet size is set to 1K bytes. The TXOP is given as 3ms (e.g., for AC_VI in 82.11e). The topology we use is the 1 1 grid as shown in Fig. 3. In order to maximize the interference between stations (and make channel switching necessary), we set stations very close, at 5M apart. We assume that the data reception range at 54Mbps rate is 8M, while the sensing range is 18M. Thus all stations are within the reception range of all other stations. Moreover, stations made backlogged with sufficient traffic. Thus they are subject to excessive contention, even if the traffic is all 1-hop traffic. Fig. 3. Simulation topology. Fig. 4 shows the result of applying Splash to the IEEE 82.11a multi-hop network. The aggregate throughput of the multi-channel 82.11a system with Splash increases from 33 to 121Mbps as the increases from 1 to 12. However, we notice that having more than 8 channels do not add to the throughput. This is because the maximum number of frames that we can transmit progressively decreases as we can see in Table I. In fact, beyond 6 channels, additional channels have only marginal value. After all, the IEEE 82.11a has been observed to support 6 channels without inter-channel interference, out of 12 orthogonal channels [1]. This effect aggregate throughput (Mbps) Fig analysis simulation Aggregate throughput in 82.11a with Splash. is negligible for small, as the fraction is relatively small with respect to the integral part of k. Fig. 5 shows the delay reduction owing to the Splash. The delay shown here is the average delay a backlogged station 2358

5 suffers to grab the channel and successfully transmit. Under the extreme contention in our simulation setting, the delay in a single-channel setting is over.5s. With Splash, however, it is decimated to 5ms at N =8channels. Again, having more than 8 channels does not help, because the additional channels are hardly utilized. delay (ms) Fig. 5. collisions Fig Average per-station inter-transmission delay in 82.11a with Splash intra-channel collisions multi-channel collisions Collisions in 82.11a with Splash. One cost we must pay by concurrently using multiple channels is the chances of multi-channel collision. Fig. 6 compares the number of intra-channel collisions and multi-channel collisions. Naturally, the multichannel collisions increase as a consequence of using more channels. This partially offsets the throughput gain from reducing the collisions within a channel. At N =12, the ratio increases to approximately 5:1. A further boosting of the performance is possible through modifying the contention window control algorithm to be less aggressive [11]. Namely, we set CW = CW/2 upon successful transmission instead of resetting it to CW min.fig. 7 shows the impact of this modification. We notice that the absolute number of collisions significantly decreased. The multi-channel collisions decreased by 45%, whereas the intrachannel collisions was more drastically reduced by 72%. The consequence of the reduced contention leads to the improvement of throughput. Fig. 8 shows that another 4Mbps can be added so that close to 16Mbps throughput is possible under the Splash scheme. collisions Fig. 7. aggregate throughput (Mbps) intra-channel collisions multi-channel collisions Collisions under the modified backoff, with Splash backoff modified backoff Fig. 8. Throughput improvement due to backoff algorithm modification under Splash. IV. CONCLUSION This paper discusses a simple channel migration scheme for multi-channel multi-hop wireless networks based on the IEEE technology. The core idea of the Splash scheme is simply that whenever a pair of stations starts communication on a channel, all other stations jump to another channel next in priority. Simulation and analysis show that this simple scheme boosts the throughput by 4 to 5 times, with tantamount reduction in delay. In our future work, we will investigate the effect of allowing the stations in the migrated channels to return to the default channel after they finish their TXOP. Also, we will see if a cross-layer assistance for an early jump will enable Splash to work for single frame exchanges as well. Thirdly, we will attempt to determine a time threshold over which Splash is triggered, in order to prevent channel migration overhead when bursts are short. REFERENCES [1] Hyogon Kim, Kyuyoung Choi, Heejo Lee and Inhye Kang, A simple congestion-resilient link adaptation algorithm for IEEE WLANs, IEEE Global Telecommunications Confernce (Globecom), November 26 [2] S. L. Wu, C. Y. Lin, Y. C. Tseng, and J. P. Sheu, A New Multi- Channel MAC Protocol with On-Demand Channel Assignment for Multi-Hop Mobile Ad Hoc Networks, International Symposium on Parallel Architectures, Algorithms and Networks (I-SPAN), December

6 [3] J. So and N. Vaidya, Multi-Channel MAC for Ad Hoc Networks: Handling Multi-Channel Hidden Terminals Using A Single Transceiver, 5th International Symposium Mobile Ad Hoc Networking and Computing (MobiHoc), pp , May 24. [4] P. Bahi, R. Chandra, and J. Dunagan, SSCH: Slotted Seeded Channel Hopping for Capacity Improvement in IEEE Wireless Networks, 1th International Conference on Mobile Computing and Networking (MobiCom), pp , Oct. 24. [5] IEEE P82.11n/D2., Draft STANDARD for Information Technology- Telecommunications and information exchange between systems- Local and metropolitan area networks- Specific requirements- Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications, Feb. 27. [6] ANSI/IEEE, 82.11e: Wireless Medium Access Control (MAC) and Physical Layer (PHY) Specifications: Medium Access Control (MAC) Enhancements for Quality of Serivce (QoS), Nov. 22. [7] Bianchi, G., Performance analysis of the IEEE distributed coordinationfunction, IEEE Journal on Selected Areas in Communications, Volume 18, Issue 3, Mar 2 Page(s): [8] Maxim 2.4GHz 82.11b Zero-IF Transceivers. [9] F. Herzel, G. Fischer, and H. Gustat, An Integrated CMOS RF Synthesizer for 82.11a Wireless LAN, IEEE Journal of Solid-state Circuits, 18(1), October 23. [1] Yunxin Liu, Yongqiang Xiong, Yang Yang, Pengzhi Xu, and Qian Zhang, An Experimental Study on Multi-channel Multi-radio Multihop Wireless Networks, IEEE Globecom 25. [11] I. N. Vukovic, N. Smavatkul, Saturation throughput analysis of different backoff algorithms in IEEE82.11, in proceedings of IEEE PIMRC