Link-level Measurements from an 802.11b Mesh NetworkDaniel Aguayo John Bicket Sanjit Biswas Glenn Judd†Robert MorrisM.I.T. Computer Science and Artificial Intelligence Laboratory{aguayo, jbicket, biswas, rtm}@csail.mit.edu†Carnegie Mellon [email protected] paper analyzes the causes of packet loss in a 38-nodeurban multi-hop 802.11b network. The patterns and causesof loss are important in the design of routing and error-correction protocols, as well as in network planning.The paper makes the following observations. The distri-bution of inter-node loss rates is relatively uniform over thewhole range of loss rates; there is no clear threshold sepa-rating “in range” and “out of range.” Most links have rela-tively stable loss rates from one second to the next, thougha small minority have very bursty losses at that time scale.Signal-to-noise ratio and distance have little predictive valuefor loss rate. The large number of links with intermediateloss rates is probably due to multi-path fading rather thanattenuation or interference.The phenomena discussed here are all well-known. Thecontributions of this paper are an understanding of theirrelative importance, of how they interact, and of the impli-cations for MAC and routing protocol design.Categories and Subject DescriptorsC.2.1 [Computer Communication Networks]: NetworkArchitecture and Design—Wireless communicationGeneral TermsMeasurement,PerformanceKeywordswireless, mesh, 802.11b1. IntroductionThis paper is a measurement study of the Roofnet multi-hop wireless network. Roofnet nodes are computers withPermission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.SIGCOMM’04, Aug. 30–Sept. 3, 2004, Portland, Oregon, USA.Copyright 2004 ACM 1-58113-862-8/04/0008 ...$5.00.802.11b cards in apartments spread over six square kilome-ters of Cambridge, Massachusetts. Each node has a roof-mounted omni-directional antenna. The network’s mainpurpose is to provide Internet access via a few wired gate-ways. The initial implementation strategy was to combineexisting radio, MAC, and routing technology in order tobuild a production-quality network as quickly as possible.This approach led to performance far less than expected,primarily due to assumptions made by MAC and routingprotocols that were a poor fit to the network’s actual be-havior. It is widely understood that wireless differs fromsimple abstract models in a number of ways [10]; the goalof this paper is to provide insight into which differences areimportant enough to worry about, and to draw conclusionsrelevant to the design of future MAC and routing protocols.Many routing and link-layer protocols assume the validityof a “neighbor” abstraction that partitions all the pairs ofnodes into pairs that can communicate directly, and pairsthat cannot. This assumption justifies the use of graph-theoretic routing algorithms borrowed from wired networks,where the assumption is true. It leads to the design of MACprotocols such as 802.11 that assume that a pair of nodes willeither hear each other’s control packets (e.g. RTS/CTS),or will not interfere. It justifies conservative transmit bit-rate selection algorithms that reduce the bit-rate after a fewpacket losses. Many existing protocols might have to be re-designed if the neighbor abstraction turned out to be a poorapproximation of reality.In principle the neighbor abstraction is supported by typ-ical assumptions about the relationship between signal-to-noise ratio and bit error rate (S/N and BER). This relation-ship is typically assumed to have a rapid transition fromessentially zero BER to a BER high enough to corrupt ev-ery packet. For example, the transition zone for the IntersilPrism HFA3873 baseband processor is about 3 dB, regard-less of bit rate [1]. Since signal strength falls off rapidly withdistance, one might expect relatively few node pairs to liein the transition zone. As a result, one might expect almostall pairs of nodes to either be able to talk to each other withlow loss, or not at all. Some empirical 802.11 measurementssuggest that the neighbor abstraction usually holds [7, 10],while others do not [6, 11].This paper starts with the observation that most Roofnetnode pairs that can communicate at all have intermediateloss rates; that is, the neighbor abstraction is a poor approx-Links with intermediate loss rates are com-mon, with no sharp transition between highand low packet loss rates.Sec. 3Inter-node distance is not strongly correlatedwith whether nodes can communicate.Sec. 4Most links have non-bursty loss patterns. Sec. 5Links with very high signal strengths arelikely to have low loss rates, but in generalsignal strength has little predictive value.Sec. 6A link is likely to have a significant loss rateat its optimum 802.11b bit-rate.Sec. 7Multi-path fading greatly affects outdoorlinks and helps explain intermediate lossrates.Sec. 9Figure 1: Summary of major conclusions for wirelessMAC and routing protocol design.Figure 2: A map of Roofnet, with a black dot foreach of the 38 nodes that participated in the exper-iments presented in this paper.imation of reality. The remainder of the paper explores aseries of hypotheses for the causes of packet loss in Roofnet,and for the predominance of intermediate loss rates. Thehypotheses include factors that affect signal-to-noise ratio(distance and interference), choice of transmit bit rate, andmulti-path fading. Figure 1 lists the paper’s main conclu-sions about these sources of packet loss. The conclusions inthis paper should not be viewed as universal, since they arelimited by the particulars of Roofnet’s configuration.2. Experimental MethodologyRoofnet consists of 38 nodes distributed over roughly sixsquare kilometers of Cambridge. Each consists of a PC withan 802.11b card connected to an omni-directional antennamounted on the roof. Figure 2 shows a map of the network.The area is dominated by tightly-packed three- and four-story houses; most antennas are mounted about two or threefeet above the chimneys of these houses. There are also anumber of taller buildings in the area; seven Roofnet