Estimation of Link Interference in Static Multi-hop Wireless ..., Lecture notes of Wireless Networking

Download Estimation of Link Interference in Static Multi-hop Wireless ... and more Lecture notes Wireless Networking in PDF only on Docsity! Estimation of Link Interference in Static Multi-hop Wireless Networks Jitendra Padhye†, Sharad Agarwal†, Venkata N. Padmanabhan†, Lili Qiu‡, Ananth Rao§, Brian Zill† †Microsoft Research ‡University of Texas, Austin §University of California, Berkeley We present a measurement-based study of interference among links in a static, IEEE 802.11, multi-hop wireless network. In- terference is a key cause of performance degradation in such net- works. To improve, or to even estimate the performance of these networks, one must have some knowledge of which links in the network interfere with one another, and to what extent. However, the problem of estimating the interference among links of a multi- hop wireless network is a challenging one. Accurate modeling of radio signal propagation is difficult since many environment and hardware-specific factors must be considered. Empirically testing every group of links is not practical: a network with n nodes can have O(n2) links, and even if we consider only pairwise interfer- ence, we may have to potentially test O(n4) pairs. Given these difficulties, much of the previous work on wireless networks has assumed that information about interference in the network is ei- ther known, or that it can be approximated using simple heuristics. We test these heuristics in our testbed and find them to be inaccu- rate. We then propose a simple, empirical estimation methodology that can predict pairwise interference using only O(n2) measure- ments. Our methodology is applicable to any wireless network that uses omni-directional antennas. The predictions made by our methodology match well with the observed pairwise interference among links in our 22 node, 802.11-based testbed. 1 Introduction Multi-hop wireless networks have been a subject of much study. Most of the original work in this area was moti- vated by scenarios in which the nodes were highly mo- bile. Recently, interesting commercial applications of static multi-hop wireless networks have emerged. One example of such applications is “community wireless networks” [2, 11]. Several companies [9, 14] are field-testing wireless “mesh” networks to provide broadband Internet access. Interference among wireless links significantly impacts the performance of static multi-hop wireless networks. Sev- eral researchers have studied this issue. The impact of inter- ference on the capacity of wireless networks is studied in [8, 10, 12], while the impact on the performance of transport- level protocols is considered in [6, 7, 13]. The need for rout- ing protocols to take link interference into account has been underscored in [4, 5]. Information about link interference is also needed for optimal channel assignment [16]. Many of these studies build upon the knowledge of which links in the network interfere with each other. Yet, the prob- lem of estimating the interference among links in a multi- hop wireless network has not been adequately addressed. The problem of estimating link interference can be de- scribed informally as follows: given a set of wireless links, estimate whether (and by how much) their aggerate throughput will decrease when the links are all active simul- taneously, compared to when they are active individually. This is a challenging problem for several reasons. Accu- rate modeling of radio signal propagation is difficult since many environment and hardware-specific factors must be considered. Empirically testing every group of links for interference is not practical: a network with n nodes can have O(n2) links, and even if we consider only pairwise interference, we may potentially have to test O(n4) pairs. The interference pattern could change due to environmen- tal factors, so interference estimation is not a one-time task. Hence, it is important to do it efficiently. Given these difficulties, some researchers simply assume that the information about which links in the network inter- fere with each other is known apriori [10]. Others assume that it can be approximated by using simple heuristics. One common heuristic states that the interference range equals a small multiple (typically, a factor of 2) of the communica- tion distance [8, 18]. In this paper, we study the phenomenon of interference among links in a 22-node IEEE 802.11 a/b/g based indoor wireless testbed. The paper makes two contributions. First, we show that the simple heuristics described in the previous literature fail to accurately predict the interference among links in our testbed. Second, we propose a simple, empiri- cal estimation methodology to predict pairwise interference that requires only O(n2) measurement experiments. We show that the predictions made by our methodology match well with the observed pairwise interference among links in our testbed under a variety of conditions. Our methodology is useful for any wireless network where nodes use omni- directional antennas. We focus on omni-directional anten- nas since these are cheap and easy to deploy, and hence popular. Network architectures based on omni-directional antennas are quite common [17]. Paper outline: First, we formalize the notion of pair- wise interference among wireless links. Next, we present a brief description of our testbed. We show that the simple heuristics used in previous work do not accurately model the interference in our testbed network. We next present our empirical methodology, and show that it accurately pre- dicts pairwise link interference in our testbed. Finally, we summarize related work and present our conclusions. 2 Interference among wireless links In this section, we define a metric to measure interference between a pair of wireless links. We assume that nodes communicate using the IEEE 802.11 protocol; parameters such as transmit power, data rate etc. are all set to fixed values; and the background noise level is constant. We also assume that RTS/CTS handshake is disabled for all nodes, which is the default behavior for most wireless cards. We start by defining what constitutes a wireless link. Un- like in a wired network, the links in a wireless network are not well-defined. For the purposes of this paper, we define wireless links using packet loss rate. We say that a link from node A to node B, denoted by LAB , exists if the packet loss rate in either direction does not exceed some thresh- old. We defer a detailed discussion of the definition until Section 4.2. We now define a metric to measure interference between a pair of links. Consider links LAB and LCD. For some fixed packet size, let UAB denote the unicast throughput of the link LAB , when no other links are active in the net- work. Similarly define UCD for link LCD. Now assume that both LAB and LCD are active simultaneously. Let their respective unicast throughput be denoted by U AB,CD AB and UAB,CD CD . Define the link interference ratio as: LIRAB,CD = UAB,CD AB + UAB,CD CD UAB + UCD (1) Thus, LIR is the ratio of aggregate throughput of the links when they are active simultaneously, to their aggregate throughput when they active individually. LIR takes values between 1 and 0. The maximum value of LIR is 1, which means that the aggregate throughput does not decrease when the links are active simultaneously. Thus, LIR = 1 implies that the links do not interfere. A value of LIR less than 1 means that the aggregate through- put of the links decreases when they operate simultane- ously. Thus, LIR < 1 implies that the links interfere with each other. The links can interfere with each other due to several reasons, listed below. Consider two links, LAB and LCD: Carrier Sense: The 802.11 protocol requires the sender to monitor the radio channel for signs of activity, prior to transmitting a packet. If any activity is detected, transmis- sion is deferred until a later time 1. This is known as carrier sensing. If the two senders, A and C are within the carrier sense range of each other, then only one of them will trans- mit at a time. Otherwise, they may both transmit, and one of the following may occur. Data-Data Collision: The transmission by C may gen- erate sufficient noise at B to interfere with reception of the packet being sent by A. A similar “collision” may occur at D. This is known as the hidden terminal problem. Data-ACK Collision: For unicast communication, the 802.11 protocol requires the receiver of a packet to transmit an acknowledgment to the sender. If node D successfully receives the data packet sent by C, it will transmit an ACK. This transmission may interfere with ongoing reception of data packet at B. A similar collision may occur at D. ACK-Data Collision: The data packet sent by C may interfere with ongoing reception of ACK sent by B at A. A similar collision may occur at C. ACK-ACK Collision: The ACK sent by D may inter- fere with the reception of ACK sent by B at A. A similar collision may occur at C. 1This is a simplified description of the actual protocol. 26 24 25 19 27 15 17 20 18 16 11 14 09 04 01 02 06 05 08 0307 10 Figure 1: Layout of our testbed A typical value of LIR is 0.5, which means that the ag- gregate throughput of the links is halved when they are ac- tive together. This usually (but not always) happens when the senders are within carrier sense range of each other. The minimum value of LIR is 0. This means that the links get zero throughput when they operate together. This can hap- pen if the senders are not within the carrier sense range of each other, and collisions at the receiver are frequent. In practice, we see a range of LIR values instead of just the three described above. They can result from packet losses, variable nature of background noise etc. Many of the simple heuristics used to estimate link interference only predict whether a pair of links interfere with each other or not. In other words, they only predict LIR is less than 1. We term this as the “binary” notion of interference. 3 Testbed The experimental data reported in this paper was collected using a 22-node wireless testbed, located on one floor of a typical office building. The nodes are placed in offices, conference rooms and labs (Figure 1). All rooms have floor-to-ceiling walls and wooden doors. The nodes were not moved during testing. Each node is equipped with two 802.11 wireless cards: a Proxim ORiNOCO a/b/g Combo- Card Gold, and a NetGear a/b/g WAG 511. For experiments described in this paper, the two cards were never active si- multaneously. RTS/CTS handshake was disabled (by de- fault). All cards operated in the 802.11 ad-hoc mode. We used the built-in antennas of these cards, which are roughly omni-directional. 4 Performance of Simple Heuristics In this section, we consider the performance of three sim- ple heuristics from previous literature. The experiments described in this section use the following settings. All Orinoco cards were turned off. All Netgear cards were set to operate in 802.11a mode, on channel 36, at full transmit power. The transmit rate of each card was fixed at 6Mbps. We use 802.11a mode for these tests since our building has an operational 802.11b network. 4.1 The Heuristics The first heuristic that we consider is used in [4, 5]. It as- sumes that all links on a multi-hop wireless path interfere with each other. In a connected network, we can always construct a path that includes a given pair of links. In short, this heuristic assumes that any two links in the network in- terfere with each other. This is clearly a pessimistic model. Thus, a lost unicast packet has a higher impact on through- put measured at the user level. Second, while ACK packets may not collide with one another, data and ACK packets can still collide. We now test our hypothesis experimentally. 5.1 Evaluation: Baseline scenario To test the hypothesis, we performed the following exper- iment. We use the same settings (802.11a, full transmit power, transmission rate fixed at 6Mbps) that we used in Section 4 and consider the same 75 link pairs as shown in Figure 2. To minimize the impact of environmental factors, the broadcast experiments designed to measure BIR were performed just before the unicast experiments designed to measure LIR. The median values of BIR and LIR for each link pair are shown in Figure 3. We see that BIR matches LIR well in most cases. The CDF of the absolute error (|LIR − BIR|) is shown in Figure 4. The median of absolute error is zero, and the mean is 0.026. Given that |LIR−BIR| can range from 0 to 1, the mean and the me- dian are quite low. Thus, our methodology works quite well in this scenario. These results bring up several interesting questions. First, does the methodology work for other scenarios? Second, note that we carried out the broadcast and the unicast exper- iments back-to-back. In reality, we must do all the broad- cast experiments together, and then use the results to pre- dict link interference. The question then becomes: if we do broadcast experiments separately, will BIR obtained at some point in time still match LIR observed at some later point? Third, is the model capable of telling us why two links interfere? We discuss these questions next. 5.2 Other scenarios We considered three other scenarios to evaluate our method- ology. In the first scenario, we turned on the autorate feature for each card. When the autorate algorithm is on, the trans- mission rate for unicast packets may vary over time, in re- sponse to changing noise levels etc. The rate selection algo- rithm is not standardized. The broadcast packets, however, are always sent at the lowest data rate (6Mbps for 802.11a). Note that in the baseline scenario, the unicast transmission rate was also fixed at 6Mbps. With autorate on, we would expect more mismatch between BIR and LIR. In the second scenario, we reduced the transmit power on each card to 50% of the full power. We fixed the transmis- sion rate at 6Mbps. At 50% transmit power, the network has fewer links: only 128, instead of 152. Thus, the link pairs used in this scenario are different from the link pairs used in the previous, full-power scenarios. The average loss rate of these 128 links is 4.6%, while the average loss rate at full power was 2.9%. Since, the links are more lossy in this sce- nario, we would expect slightly higher mismatch between BIR and LIR in this setting. All the experiments so far were done in 802.11a mode, using the NetGear cards. In the third scenario, we turned off the Netgear cards, and used Orinoco cards, set to op- erate in 802.11g mode (i.e. in 2.4GHz spectrum), at full power, with rate fixed at 1Mbps (i.e. the lowest data rate for 802.11g). We have an infrastructure mode 802.11b net- work in our building, which operates in the same frequency band. We tested this scenario at night, to minimize the im- pact of interference from the WLAN, however, we would still expect to see higher error in this scenario. For each of these scenarios, we measured BIR and LIR of 75 link pairs, using back-to-back experiments as before. The CDF of absolute error in each of the three cases is shown in Figure 5. The results show that our methodol- ogy performs generally well in each scenario. As expected, the mismatch is somewhat high for the autorate scenario. In the other two cases, the median error is only 0.01. Even in autorate case the median error is only 0.03, while the mean is 0.065. These three experiments increase our confidence in the general applicability of our method. 5.3 BIR and LIR measured 5 days apart We used the same settings as the baseline scenario, but did only the broadcast experiments. We compare BIR calcu- lated from these experiments with the LIR measured in the baseline experiment. The two experiments were done 5 days apart. The CDF of absolute error is shown in Fig- ure 6. The graph also shows the baseline error CDF (la- beled “Back-to-back”’) for comparison purposes. We see that BIR is still generally a good predictor of LIR, but as expected, the error is somewhat higher compared to the baseline (back-to-back) case. The median error is only 0.01, and mean is 0.049. The results show that even in a static environment like ours, the interference patterns are slightly different at different times. The need to repeat interference measurements underscores the need for an inexpensive ex- perimental methodology to measure interference. 5.4 Why do links interfere? Our methodology also helps determine why two links inter- fere with one another. Consider links LAB and LCD that in- terfere with one another. During the broadcast experiments done to determine BIR, we had nodes A and C broadcast alone, as well as together. Consider the ratio of their send rates, when they were broadcasting together to when they were broadcasting alone. Define carrier sense ratio: CSR = (SAC A + SAC C )/(SA + SC). (3) Note that we are using broadcast packets, so both senders send at the same data rate. If two senders are within the carrier sense range of each other, then only one of them would be able to send at a time, resulting in a CSR value of 0.5. If the senders are not within each other’s carrier sense range, CSR will be 1. Intermediate values can result from noise, differences in sensitivity of antennas, signal strength fluctuations due to environmental factors etc. In the baseline scenario shown in Figure 2, 46 link pairs have LIR < 0.9, indicating some degree of interference. Of these, 34 link pairs had a CSR of 0.5. Thus, carrier sensing seems to be the major cause of interference in our testbed. We see similar results for the other three scenarios considered in Section 5.2. We believe that this is one of the reasons why BIR and LIR show a good match under all scenarios. We are currently investigating this issue further. 0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 F ra ct io n abs(LIR-BIR) Figure 4: Baseline Scenario 0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 F ra ct io n abs(LIR-BIR) 11a, 50% Power, Fix Rate 11g, Full Power, Fix Rate. 11a, Full Power, AutoRate. Figure 5: Three other scenarios. 0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 F ra ct io n abs(LIR-BIR) 5 Days Apart Back-to-back Figure 6: Measured 5 days apart 6 Related Work The importance of studying wireless interference has long been recognized. For example, in [15] the impact of inter- ference on fairness is considered. In [19] it is shown that paths with high degree of interference suffer disproportion- ately. Several researchers [8, 10, 12, 13] have considered impact of interference on the overall capacity of a multi-hop wireless network. However, each of these papers assumes that the information about link interference is available, but they do not describe how to estimate it. Thus our work on a practical method for estimating link interference helps gen- erate the information taken for granted in previous work. As we discussed earlier, various heuristics for estimating link interference have been proposed [4, 8, 12, 18]. We have shown that our empirical methodology can provide a more accurate estimation of pairwise link interference. The knowledge of which links interfere with one another can benefit a number of network operations. For example, it can improve routing algorithms [4, 5], help in engineering and managing multi-hop wireless networks [3], and aid in channel assignment [16]. There is a large body of work on measuring various prop- erties of wireless networks. Here we list some of the recent work. Aguayo et al. [1] analyze the causes of packet loss in an outdoor multihop 802.11b network. Yarvis et al [20] use testbeds in three different houses to study the properties of home wireless networks. Gupta et al [7] experimentally study the performance of TCP in a multi-hop wireless net- work. Our work contributes a practical technique for mea- suring another key property, viz., wireless interference. 7 Conclusion and Ongoing Work In this paper, we considered the problem of estimating pairwise interference among links in a multi-hop wireless testbed. Using experiments done in a 22-node, 802.11- based testbed, we showed that some of the previously- proposed heuristics for predicting pairwise interference are inaccurate. We then proposed a simple, empirical method- ology to estimate pairwise interference using only O(n2) measurements. We showed that our methodology accu- rately predicts pairwise interference among links in our testbed in a variety of settings. Our methodology is applica- ble to any 802.11-based wireless network where nodes use omni-directional antennas. There are several avenues for future work. We hope to increase the accuracy of our methodology by accounting for the impact of four factors that we ignored in this paper. These four factors are: (i) retransmissions of lost unicast packets, (ii) RTS/CTS handshake (iii) collisions between data and ACK packets (iv) autorate algorithms. We would like to extend our approach to estimate inter- ference among larger groups of links, instead of just pair- wise interference. Finally, we note that our methodology requires nodes to generate broadcast traffic, and existing traffic on the net- work can significantly reduce the accuracy of our approach. We are currently exploring the possibility of determining in- terference patterns by simply observing correlation between existing traffic flows on the network. References [1] D. Aguago, J. Bicket, S. Biswas, G. Judd, and R. Morris. Link-level measurements from an 802.11b mesh network. In SIGCOMM, 2004. [2] Bay area wireless users group. http://www.bawug.org/. [3] R. Chandra, L. Qiu, K. Jain, and M. Mahdian. On the placement of internet taps in wireless neighborhood networks. In ICNP, 2004. [4] D. De Couto, D. Aguayo, J. Bicket, and R. Morris. High-throughput path metric for multi-hop wireless routing. In MOBICOM, 2003. [5] R. Draves, J. Padhye, and B. Zill. Routing in multi-radio, multi-hop wireless mesh network. In MOBICOM, 2004. [6] Z. Fu, P. Zerfos, H. Luo, S. Lu, L. Zhang, and M. Gerla. The Impact of Multihop Wireless Channel on TCP Throughput and Loss. In INFOCOM, 2003. [7] A. Gupta, I. Wormsbecker, and C. Williamson. Experimental evalu- ation of TCP performance in multi-hop wireless ad hoc networks. In MASCOTS, 2004. [8] P. Gupta and P. R. Kumar. The capacity of wireless networks. IEEE Trans on Info Theory, Mar 2000. [9] Invisible Networks. http://www.invisible.uk.net/how/. [10] K. Jain, J. Padhye, V. Padmanabhan, and L. Qiu. The impact of inter- ference on multi-hop wireless network performance. In MOBICOM, 2003. [11] R. Karrer, A. Sabharwal, and E. Knightly. Enabling Large-scale Wireless Broadband: The Case for TAPs. In HotNets, 2003. [12] M. Kodialam and T. Nandagopal. Charaterizing achievable rates in multi-hop wireless newtorks: The joint routing and scheduling prob- lem. In MOBICOM, Sep. 2003. [13] J. Li, C. Blake, D. S. J. De Couto, H. I. Lee, and R. Morris. Capacity of ad hoc wireless networks. In MOBICOM, 2001. [14] Mesh Networks Inc. http://www.meshnetworks.com. [15] T. Nandagopal, T. Kim, X. Gao, and V. Bharghavan. Achieving MAC Layer Fairness in wireless packet networks. In MOBICOM, 2000. [16] A. Rainwala and T. Chiueh. Architecture and algorithms for an IEEE 802.11-based multi-channel wireless mesh network. In INFOCOM, 2005. [17] MIT roofnet. http://www.pdos.lcs.mit.edu/roofnet/. [18] K. Xu, M. Gerla, and S. Bae. How effective is the IEEE 802.11 RTS/CTS handshake in ad hoc networks? In GLOBECOM, 2002. [19] X. Yang and N. H. Vaidya. Priority scheduling in wireless ad hoc networks. In MOBIHOC, June 2002. [20] M. Yarvis, K. Papagiannaki, and W. S. Conner. Characterization of 802.11 wireless networks in the home. In WiNMeE, 2005.