This file is raw output from pdftotext and may not be ideal for distribution. If you are a maintainer for Hackipedia, please sit down when you have time and clean this text version up. Source PDF: /mnt/fw-js/docs/Hardware/Zigbee/A Comprehensive Study Of Ieee 802.15.4 And Zigbee 506.pdf Like all conversions the text below should be fully readable as UTF-8 unicode text. --------------------------------------------------------------- 1 A Comprehensive Performance Study of IEEE 802.15.4 Jianliang Zheng and Myung J. Lee Abstract— IEEE 802.15.4 is a new standard uniquely network more cost-efficient than a wired network in designed for low rate wireless personal area networks (LR- general. WPANs). It targets low data rate, low power consumption The release of IEEE 802.15.4 (referred to as 802.15.4 and low cost wireless networking, and offers device level wireless connectivity. We develop an NS2 simulator for hereinafter), "Wireless Medium Access Control (MAC) IEEE 802.15.4 and conduct several sets of experiments to and Physical Layer (PHY) Specifications for Low Rate study its various features, including: (1) beacon enabled Wireless Personal Area Networks (LR-WPANs)" [1]1 , mode and non-beacon enabled mode; (2) association, tree represents a milestone in wireless personal area networks formation and network auto-configuration; (3) orphaning and wireless sensor networks. 802.15.4 is a new standard and coordinator relocation; (4) carrier sense multiple access with collision avoidance (CSMA-CA), both unslotted and uniquely designed for low rate wireless personal area slotted; and (5) direct, indirect and guaranteed time slot networks. It targets low data rate, low power consump- (GTS) data transmissions. In non-beacon enabled mode tion and low cost wireless networking and offers device and under moderate data rate, the new IEEE 802.15.4 level wireless connectivity. A host of new applications standard, compared with IEEE 802.11, is more efficient can benefit from the new standard, such as those using in terms of overhead and resource consumption. It also enjoys a low hop delay (normalized by channel capacity) sensors that control lights or alarms, wall switches that on average. In beacon enabled mode, an LR-WPAN can can be moved at will, wireless computer peripherals, be flexibly configured to meet different needs, such as link controllers for interactive toys, smart tags and badges, failure self-recovery and low duty cycle. In both beacon tire pressure monitors in cars, inventory tracking devices. enabled mode and non-beacon enabled mode, association 802.15.4 distinguishes itself from other wireless stan- and tree formation proceed smoothly and the network can shape up efficiently by itself. We also discuss some issues dards such as IEEE 802.11 (referred to as 802.11 here- that could degrade the network performance if not handled inafter) [2] and Bluetooth [3] by some unique features properly. (see section II). However, there are no simulations or im- Index Terms— 802.15.4, LR-WPAN, WPAN, wireless plementations available so far to test these new features. sensor networks, low power, low data rate, (non-)beacon We develop an NS2 simulator for 802.15.4 and carry out enabled mode. several sets of experiments to evaluate its performances, in hopes of helping IEEE to verify and/or improve the design, and facilitating researchers and manufacturers to I. BACKGROUND AND M OTIVATION develop products based upon this new standard. 802.15.4 OMPARED with wired networks, wireless net- has been designed as a flexible protocol in which a C works provide advantages in deployment, cost, size, and distributed intelligence. Wireless technology set of parameters can be configured to meet different requirements. As such, we also try to find out how users not only enables users to set up a network quickly, but can tailor the protocol to their needs and where the trade- also enables them to set up a network where it is in- off is for some applications. convenient or impossible to wire cables. The “care free” The rest of the paper is structured as follows. In feature and convenience of deployment make a wireless section II, we give a brief description of 802.15.4. Next, in section III, we outline the NS2 simulator for 802.15.4. Jianliang Zheng and Myung J. Lee are with the Department of Then, in section IV, we define a set of performance Electrical Engineering, City College, The City University of New metrics and present the experimental setup. In section V, York, New York, NY 10031 USA (e-mail: zheng@ee.ccny.cuny.edu, lee@ccny.cuny.edu) we give out the experimental results with discussions. The research is supported by Samsung Advanced Institute of Tech- nology. 1 All results in this paper apply to the IEEE 802.15.4 draft D18 [1] 2 Finally, in section VI, we conclude. from MAC sublayer. The turnaround time from transmitting to receiving, or vice versa, should be II. A B RIEF D ESCRIPTION OF IEEE 802.15.4 no more than 12 symbol periods. • Energy detection (ED) within the current channel: It The new IEEE standard, 802.15.4, defines the physical is an estimate of the received signal power within layer (PHY) and medium access control sublayer (MAC) the bandwidth of an IEEE 802.15.4 channel. No specifications for low data rate wireless connectivity attempt is made to identify or decode signals on among relatively simple devices that consume minimal the channel in this procedure. The energy detection power and typically operate in the Personal Operating time shall be equal to 8 symbol periods. The result Space (POS) of 10 meters or less. An 802.15.4 net- from energy detection can be used by a network work can simply be a one-hop star, or, when lines of layer as part of a channel selection algorithm, or communication exceed 10 meters, a self-configuring, for the purpose of clear channel assessment (CCA) multi-hop network. A device in an 802.15.4 network (alone or combined with carrier sense). can use either a 64-bit IEEE address or a 16-bit short • Link quality indication (LQI) for received packets: address assigned during the association procedure, and Link quality indication measurement is performed a single 802.15.4 network can accommodate up to 64k for each received packet. The PHY layer uses re- (216 ) devices. Wireless links under 802.15.4 can operate ceiver energy detection (ED), a signal-to-noise ratio, in three license free industrial scientific medical (ISM) or a combination of these to measure the strength frequency bands. These accommodate over air data rates and/or quality of a link from which a packet is of 250 kb/sec (or expressed in symbols, 62.5 ksym/sec) received. However, the use of LQI result by the in the 2.4 GHz band, 40 kb/sec (40 ksym/sec) in the network or application layers is not specified in the 915 MHz band, and 20 kb/sec (20 ksym/sec) in the 868 standard. MHz. Total 27 channels are allocated in 802.15.4, with • Clear channel assessment (CCA) for carrier sense 16 channels in the 2.4 GHz band, 10 channels in the 915 multiple access with collision avoidance (CSMA- MHz band, and 1 channel in the 868 MHz band. CA): The PHY layer is required to perform CCA us- Wireless communications are inherently susceptible ing energy detection, carrier sense, or a combination to interception and interference. Some security research of these two. In energy detection mode, the medium has been done for WLANs and wireless sensor net- is considered busy if any energy above a predefined works [13]–[16], [20], [22], but pursuing security in energy threshold is detected. In carrier sense mode, wireless networks remains a challenging task. 802.15.4 the medium is considered busy if a signal with the employs a fully handshaked protocol for data transfer re- modulation and spreading characteristics of IEEE liability and embeds the Advanced Encryption Standard 802.15.4 is detected. And in the combined mode, (AES) [4] for secure data transfer. both conditions aforementioned need to be met in In the following subsections, we give a brief overview order to conclude that the medium is busy. of the PHY layer, MAC sublayer and some general • Channel frequency selection: Wireless links under functions of 802.15.4. Detailed information can be found 802.15.4 can operate in 27 different channels (but in [1]. a specific network can choose to support part of the channels). Hence the PHY layer should be able A. The PHY layer to tune its transceiver into a certain channel upon The PHY layer provides an interface between the receiving the request from MAC sublayer. MAC sublayer and the physical radio channel. It pro- • Data transmission and reception: This is the es- vides two services, accessed through two service access sential task of the PHY layer. Modulation and points (SAPs). These are the PHY data service and the spreading techniques are used in this part. The 2.4 PHY management service. The PHY layer is responsible GHz PHY employs a 16-ary quasi-orthogonal mod- for the following tasks: ulation technique, in which each four information • Activation and deactivation of the radio bits are mapped into a 32-chip pseudo-random noise transceiver: Turn the radio transceiver into (PN) sequence. The PN sequences for successive one of the three states, that is, transmitting, data symbols are then concatenated and modulated receiving, or off (sleeping) according to the request onto the carrier using offset quadrature phase shift 3 keying (O-QPSK). The 868/915 MHz PHY em- the active superframe to a device. These portions ploys direct sequence spread spectrum (DSSS) with are called GTSs, and comprise the contention free binary phase shift keying (BPSK) used for chip period (CFP) of the superframe. modulation and differential encoding used for data • Providing a reliable link between two peer MAC symbol encoding. Each data symbol is mapped into entities: The MAC sublayer employs various mech- a 15-chip PN sequence and the concatenated PN anisms to enhance the reliability of the link between sequences are then modulated onto the carrier using two peers, among them are the frame acknowledg- BPSK with raised cosine pulse shaping. ment and retransmission, data verification by using a 16-bit CRC, as well as CSMA-CA. B. The MAC sublayer The MAC sublayer provides an interface between the C. General Functions service specific convergence sublayer (SSCS) and the The standard gives detailed specifications of the fol- PHY layer. Like the PHY layer, the MAC sublayer also lowing items: type of device, frame structure, superframe provides two services, namely, the MAC data service structure, data transfer model, robustness, power con- and the MAC management service. The MAC sublayer sumption considerations, and security. In this subsection, is responsible for the following tasks: we give a short description of those items closely related • Generating network beacons if the device is a to our performance study, including type of device, coordinator: A coordinator can determine whether superframe structure, data transfer model, and power to work in a beacon enabled mode, in which consumption considerations. a superframe structure is used. The superframe Two different types of devices are defined in an is bounded by network beacons and divided into 802.15.4 network, a full function device (FFD) and a aNumSuperframeSlots (default value 16) equally reduced function device (RFD). An FFD can talk to sized slots. A coordinator sends out beacons pe- RFDs and other FFDs, and operate in three modes riodically to synchronize the attached devices and serving either as a PAN coordinator, a coordinator or a for other purposes (see subsection II-C). device. An RFD can only talk to an FFD and is intended • Synchronizing to the beacons: A device attached for extremely simple applications. to a coordinator operating in a beacon enabled The standard allows the optional use of a superframe mode can track the beacons to synchronize with structure. The format of the superframe is defined by the coordinator. This synchronization is important the coordinator. From Fig. 1, we can see the superframe for data polling, energy saving, and detection of comprises an active part and an optional inactive part, orphanings. and is bounded by network beacons. The length of the • Supporting personal area network (PAN) as- superframe (a.k.a. beacon interval, BI) and the length sociation and disassociation: To support self- of its active part (a.k.a. superframe duration, SD) are configuration, 802.15.4 embeds association and dis- defined as follows: BO association functions in its MAC sublayer. This not BI = aBaseSuperf rameDuration ∗ 2 SO only enables a star to be setup automatically, but SD = aBaseSuperf rameDuration ∗ 2 also allows for the creation of a self-configuring, Where, peer-to-peer network. aBaseSuperframeDuration = 960 symbols • Employing the carrier sense multiple access with BO = beacon order collision avoidance (CSMA-CA) mechanism for SO = superframe order channel access: Like most other protocols designed The values of BO and SO are determined by the coordi- for wireless networks, 802.15.4 uses CSMA-CA nator. The active part of the superframe is divided into mechanism for channel access. However, the new aNumSuperframeSlots (default value 16) equally sized standard does not include the request-to-send (RTS) slots and the beacon frame is transmitted in the first and clear-to-send (CTS) mechanism, in considera- slot of each superframe. The active part can be further tion of the low data rate used in LR-WPANs. broken down into two periods, a contention access period • Handling and maintaining the guaranteed time slot (CAP) and an optional contention free period (CFP). The (GTS) mechanism: When working in a beacon en- optional CFP may accommodate up to seven so-called abled mode, a coordinator can allocate portions of guaranteed time slots (GTSs), and a GTS may occupy 4 Beacon Beacon GTS GTS Inactive 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 CAP CFP SD = aBaseSuperframeDuration x 2SO symbols (Active) BI = aBaseSuperframeDuration x 2BO symbols Fig. 1. An Example of the Superframe Structure more than one slot period. However, a sufficient portion association response frame. Unslotted CSMA-CA of the CAP shall remain for contention based access of or slotted CSMA-CA is used in the data extraction other networked devices or new devices wishing to join procedure. the network. A slotted CSMA-CA mechanism is used • GTS data transmission: This only applies to data for channel access during the CAP. All contention based transfer between a device and its coordinator, either transactions shall be complete before the CFP begins. from the device to the coordinator or from the Also all transactions using GTSs shall be done before coordinator to the device. No CSMA-CA is needed the time of the next GTS or the end of the CFP. in GTS data transmission. Data transfer can happen in three different ways: (1) Power conservation has been one of research focuses for from a device to a coordinator; (2) from a coordinator wireless networks [9]–[12], [17], [19], [21], since most to a device; and (3) from one peer to another in a devices in wireless networks are battery powered. The peer-to-peer multi-hop network. Nevertheless, for our standard was developed with the limited power supply performance study, we classify the data transfer into the availability in mind and favors battery powered devices. following three types: The superframe structure, the indirect data transmission • Direct data transmission: This applies to all data and the BatteryLifeExtension option are all examples. If transfers, either from a device to a coordinator, the BatteryLifeExtension is set to TRUE, all contention from a coordinator to a device, or between two based transactions are required to begin within mac- peers. unslotted CSMA-CA or slotted CSMA-CA BattLifeExtPeriods (default value 6) full backoff periods is used for data transmission, depending whether after the inter-frame space (IFS) period of the beacon non-beacon enabled mode or beacon enabled mode frame. is used. • Indirect data transmission: This only applies to data III. NS2 S IMULATOR transfer from a coordinator to its devices. In this mode, a data frame is kept in a transaction list The 802.15.4 NS2 [5] simulator developed at the by the coordinator, waiting for extraction by the Joint Lab of Samsung and the City University of New corresponding device. A device can find out if it has York confirms to IEEE P802.15.4/D18 Draft. Fig. 2 a packet pending in the transaction list by checking outlines the function modules in the simulator, and a the beacon frames received from its coordinator. brief description is given below for each of the modules. Occasionally, indirect data transmission can also • Wireless Scenario Definition: It selects the rout- happen in non-beacon enabled mode. For example, ing protocol; defines the network topology; and during an association procedure, the coordinator schedules events such as initializations of PAN keeps the association response frame in its trans- coordinator, coordinators and devices, and starting action list and the device polls and extracts the (stopping) applications. It defines radio-propagation 5 model, antenna model, interface queue, traffic pat- • Hop delay: The transaction time of passing a packet tern, link error model, link and node failures, su- to a one-hop neighbor, including time of all neces- perframe structure in beacon enabled mode, radio sary processing, backoff as well as transmission, transmission range, and animation configuration. and averaged over all successful end-to-end trans- • Service Specific Convergence Sublayer (SSCS): missions within a simulation run. It is not only This is the interface between 802.15.4 MAC and used for measuring packet delivery latency, but also upper layers. It provides a way to access all the used as a negative indicator of the MAC sublayer MAC primitives, but it can also serve as a wrapper capacity. The MAC sublayer has to handle the of those primitives for convenient operations. It is packets one by one and therefore a long delay an implementation specific module and its function means a small capacity. should be tailored to the requirements of specific • RTS/CTS overhead: The ratio of request-to-send applications. (RTS) packets plus clear-to-send (CTS) packets • 802.15.4 PHY: It implements all 14 PHY primitives. sent to all the other packets sent in 802.11. This • 802.15.4 MAC: This is the main module. It imple- metric is not applicable to 802.15.4, in which ments all the 35 MAC sublayer primitives. RTS/CTS mechanism is not used. We compare the performances of 802.11 and 802.15.4 to justify the dropping of RTS/CTS mechanism in 802.15.4. Wireless • Successful association rate: The ratio of devices Scenario Definition successfully associated with a coordinator to the total devices trying to associate with a coordinator. Upper Layers In our experiments, a device will retry in one • CSMA-CA second if it fails to associate with a coordinator in Routing • Beacon and Sync. • Assoc. and Disassoc. the previous attempt. The association is considered 802.2 LLC • Direct/Indirect/GTS Tx successful if a device is able to associate with a SSCS • Filtering • ED coordinator during a simulation run, even if multiple • Error Modeling • CCA • Enhanced Nam Anima. association attempts have been made. • LQI 802.15.4 MAC • Filtering • Association efficiency: The average number of at- • Multi-Channel 802.15.4 PHY tempts per successful association. • Orphaning rate: A device is considered orphaned if NS2 it misses aMaxLostBeacons (default value 4) bea- cons from its coordinator in a row. The orphaning rate is defined as the ratio of devices orphaned at Fig. 2. NS2 Simulator for IEEE 802.15.4 least once to the total devices that are in beacon enabled mode and keep tracking beacons. This metric is not applicable to devices in non-beacon IV. P ERFORMANCE M ETRICS AND E XPERIMENTAL enabled mode or devices in beacon enabled mode S ETUP but not tracking beacons. In our experiments, all A. Performance Metrics devices in beacon enabled mode track beacons. We define the following metrics for studying the • Orphaning recovery rate: Two different versions are performance of 802.15.4. All metrics are defined with defined for this metric. One is the ratio of orphaned respect to MAC sublayer and PHY layer in order to devices that have successfully relocated their co- isolate the effects of MAC and PHY from those of upper ordinators, i.e., have recovered from orphaning, to layers. the total orphaned devices. The other is the ratio • Packet delivery ratio: The ratio of packets success- of recovered orphanings to the total orphanings, in fully received to packets sent in MAC sublayer. which multiple orphanings of a device are counted. This metric does not differentiate transmissions and No further attempt is made if the orphaning recov- retransmissions, and therefore does not reflect what ery procedure fails. percentage of upper layer payload is successfully • Collision rate: The total collisions during a simu- delivered, although they are related. lation run. 6 • Collision rate between hidden terminals: The to- is two-ray ground reflection. Beacon order (BO) and tal collisions that occur between hidden terminals superframe order (SO) take the same value in all beacon during a simulation run. Hidden terminals prevent enabled modes, that is, the optional inactive part is not carrier sense from working effectively, and therefore included in superframes. Most experiments run 10 times transmissions from them are likely to collide at with random seeds, but those with a traffic load of 0.2 a third node [23]. In 802.11, the request-to-send packet per second (pps) and those with a traffic load of (RTS) and clear-to-send (CTS) mechanism is used 0.1 pps run 20 times and 40 times respectively. Other to tackle this problem [2]. experiment specific configuration information is given • Repeated collision rate: The total collisions that in the following paragraphs corresponding to each set of happen more than once between the same pair of experiments. packets during a simulation run. • Collision distribution: The time distribution, within a superframe, of collisions. This metric is only used 88 96 92 100 64 76 68 80 72 84 61 in beacon enabled mode. 44 56 48 60 52 41 • Duty cycle: The ratio of the active duration, includ- 83 28 36 32 40 25 73 ing transmission, reception and carrier sense time, 99 51 16 20 24 13 53 85 of a transceiver to the whole session duration. 71 39 8 9 5 33 65 91 59 23 4 1 17 45 93 B. Experimental Setup 79 31 12 0 10 29 77 95 47 19 3 2 21 57 89 Five sets of experiments are designed to evaluate the 67 35 7 11 6 37 69 various performance behaviors of 802.15.4, including 87 55 15 22 18 14 49 97 those applicable to all wireless networks (such as packet 75 27 38 30 34 26 81 delivery ratio, packet delivery latency, control overhead, 43 50 58 46 54 42 and transmission collision) as well as other behaviors 63 82 70 78 66 74 62 specific to LR-WPANs (such as association, orphaning, 98 90 94 86 and different transmission methods). The first set is for non-beacon enabled mode, the second and third sets are (a) for mixed mode, that is, a combination of beacon enabled mode and non-beacon enabled mode, and the fourth and fifth sets are for beacon enabled mode. The first three sets 3 4 run in a multi-hop environment (Fig. 3 (a)), and the other two sets run in a one-hop star environment (Fig. 3 (b)). 2 0 5 Although a specific network can take a quite different topology, the two topologies used in our experiments represent the topologies currently supported by 802.15.4 1 6 and are enough for performance study purpose. General parameters: Assuming a 10−6 to 10−5 link (b) bit error rate (BER), we apply a 0.2% statistical packet error rate (PER) to all our experiments. The simulation Fig. 3. Experiment Scenarios duration is 1000 seconds, and the application traffic runs from 20 to 900 second, leaving enough time Experiment set 1 – Comparing 802.15.4 with 802.11: for the experiment to shut down gracefully. Since the The first set of experiments are used to compare the popular constant bit rate (CBR) traffic used in most performances of 802.15.4 and 802.11. Although 802.15.4 simulations is too deterministic for non-mobile wireless and Bluetooth bear more similarities from the application networks, Poisson traffic is used for all application point of view, 802.15.4 and 802.11 are more comparable sessions in our experiments. The application packet size as far as our performance study is concerned. Both is 90 bytes. Except the fifth set of experiments, all 802.15.4 and 802.11 support multi-hop network topology the other experiments use direct data transmission. The and peer-to-peer communications, which are used in radio propagation model adopted in all our experiments our first set of experiments. The dominant topology in 7 Bluetooth, on the other hand, is one-hop star or so-called coordinators and different beacon orders. The same piconet, which consists of one coordinator and up to network topology, transmission range, frequency band, seven devices. In a piconet, a device only communicates data rate, and peer-to-peer application sessions are used with its coordinator. Although scatternets can be used as in the first set of experiments. Except node 0, which is to extend the coverage and the number of devices of a the PAN coordinator, and the leaf nodes depicted in grey, Bluetooth network, our research work showed that there which are pure devices, all the other nodes serve as both are scalability problems in scatternets [7]. Furthermore, a coordinator (to its children) and a device (to its parent). all the devices in either 802.15.4 or 802.11 share a single So we have 73 coordinators and 100 devices. This set chip code for spread spectrum, while different devices of experiments run in a mixed mode, with different in Bluetooth are assigned different chip codes. Based on percentage of coordinators beaconing (0%, 25%, 50%, the above facts, we select 802.11 instead of Bluetooth for 75% and 100%). The beacon order varies and takes the comparison. The performance is evaluated with respect values of 0, 1, 2, 3, 4, 5, 6 and 10. The application traffic to the following parameters as well as those listed in the is fixed at 1 pps. previous paragraph: Experiment set 3 – Orphaning: The third set of 2 • 101 nodes evenly distributed in an 80 x 80 m area experiments are used to study the device orphaning (Fig. 3 (a)). behavior, namely, how often orphanings happen and • 9 meter transmission range, which only covers the what percentage of orphanings, in terms of number neighbors along diagonal direction. of orphaned devices or number of orphanings, can be • 802.15.4 operates at an over air data rate of 250 recovered. The experimental setup is the same as that of kbps (in the 2.4 GHz ISM band) and in non-beacon the second set of experiments. enabled mode, and 802.11 operates at a data rate of Experiment set 4 – Collision: The fourth set of ex- 2 Mbps. periments target the collision behavior of 802.15.4. The • Poisson traffic with the following average packet experiments run in a beacon enabled star environment. rates: 0.1 packet per second (pps), 0.2 pps, 1 pps, Nevertheless, except some beacon specific metrics, most 5 pps and 10 pps. of the metrics extracted from this set of experiments • We apply two types of application traffic: (1) peer- are general and can serve for both beacon and non- to-peer application traffic, which consists of six beacon enabled modes. Besides the general parameters application sessions between the following nodes: given above, the following parameters are used in the 64 → 62, 63 → 61, 99 → 85, 87 → 97, 88 → 98, experiments: and 100 → 86, and (2) multiple-to-one application • 7 nodes form a star with a radius of 10 meters, with traffic, which consists of twelve application sessions one coordinator at the center and six devices evenly from nodes 64, 62, 63, 61, 99, 85, 87, 97, 88, 98, distributed around it (Fig. 3 (b)). 100 and 86 to node 0. The first type of application • 15 meter transmission range, which enables the traffic is used to study the general peer-to-peer be- coordinator to reach all the devices. However, a havior of 802.15.4 and, for comparison, it is applied device can only reach the coordinator and two to both 802.15.4 and 802.11. The second type of devices adjacent to it. In other words, devices are application traffic targets the important application hidden from each other unless they are adjacent to of 802.15.4, wireless sensor networks, where traffic each other. is typically between multiple source nodes and a • Operates at an over air data rate of 250 kbps (in the sink. It is only applied to 802.15.4. Although the 2.4 GHz ISM band). second type of application traffic is not used for • Poisson traffic with the average packet rate of 1 pps. comparing 802.15.4 with 802.11, we include it here • Six application sessions, one for each device, are to facilitate the comparison of 802.15.4 behaviors setup from the devices to the coordinator. under different application traffic. We refer to the • The beacon order changes from 0 to 8. second type of application traffic as sink-type ap- Experiment set 5 – Direct, indirect and GTS data plication traffic hereinafter. transmissions: The last set of experiments are used to Experiment set 2 – Association efficiency: The second investigate the different features of the three data trans- set of experiments are designed to evaluate the asso- mission methods in 802.15.4. We compare the packet ciation efficiency under different number of beaconing delivery ratio, hop delay and duty cycle of the three 8 different methods. All the parameters are the same as 802.15.4s to denote the data series corresponding to peer- those in the fourth set of experiments, except that only to-peer application traffic and sink-type application traf- two application sessions originating from adjacent de- fic respectively (see Fig. 4 and Fig. 6). However, when vices are used, and that three different data transmission experiment results are not specific to a certain application methods are used. traffic (e.g., the data series 802.15.4 in Fig. 5) or only one application traffic is applied (e.g. for 802.11), the V. E XPERIMENTAL R ESULTS protocol name is used only to denote the corresponding A. Comparing IEEE 802.15.4 with IEEE 802.11 data series. For peer-to-peer application traffic, as shown in Fig. 4, the packet delivery ratio of 802.11 decreases slowly from 99.53% to 98.65% when the traffic load changes from 100 Packet Delivery 90 0.1 packet per second (pps) to 10 pps. On the other Ratio (%) 80 802.11 hand, the packet delivery ratio of 802.15.4 drops from 70 802.15.4p 60 98.51% to 78.26% for the same traffic load change (data 802.15.4s 50 series 802.15.4p in Fig. 4). For sink-type application 40 0.1 0.2 1.0 5.0 10.0 traffic, the packet delivery ratio of 802.15.4 drops more Traffic Load (pkts/sec) sharply from 95.40% to 55.26% when the traffic load changes from 0.1 pps to 10 pps (data series 802.15.4s in Fig. 4). In general, 802.15.4 maintains a high packet Fig. 4. Comparing 802.15.4 with 802.11: Packet Delivery Ratio delivery ratio for application traffic up to 1 pps (95.70% for 802.15.4p and 87.58% for 802.15.4s), but the value decreases quickly as traffic load increases. The difference of packet delivery ratio between 802.15.4 and 802.11 comes from the fact that the former (RTS + CTS) Pkts 3 per Poisson Pkt does not use RTS/CTS mechanism while the latter does. 2 802.11 This RTS/CTS overhead proves to be useful when traffic 1 802.15.4 load is high, but obviously too expensive for low data 0 rate applications as of the case of LR-WPANs for which 0.1 0.2 1.0 5.0 10.0 802.15.4 is designed. From Fig. 5, we can see the Traffic Load (pkts/sec) ratio of (RTS+CTS) packets to Poisson data packets is within the scope [2.02, 2.78], which cannot be justified in 802.15.4, considering the less than 4% increase of Fig. 5. Comparing 802.15.4 with 802.11: RTS /CTS Overhead packet delivery ratio for application traffic up to 1 pps. Note that, even under collision-free condition, the ratio of (RTS+CTS) packets to Poisson data packets is larger than 2.0, because RTS/CTS packets are also used 0.025 for transmissions of other control packets such AODV Hop Delay (sec) 0.020 802.11 packets. It is clear that the high ratio of (RTS+CTS) 0.015 802.11* packets to Poisson data packets for 0.1 pps must come 0.010 802.15.4p from the high ratio of other control packets to Poisson 0.005 802.15.4s 0.000 data packets, since collisions are ignorable under such 0.1 0.2 1.0 5.0 10.0 low traffic load. Traffic Load (pkts/sec) The RTS/CTS mechanism also affects the network latency. We measure the average hop delay for both protocols in comparison, and the results are depicted Fig. 6. Comparing 802.15.4 and 802.11: Hop Delay in Fig. 6. The initial results show that 802.11 enjoys a lower delay than 802.15.4 (data series 802.11 and To distinguish experiment results for 802.15.4 with 802.15.4p in Fig. 6). Nevertheless, this comparison is different application traffic, we use 802.15.4p and unfair to 802.15.4, since it operates at a data rate of 9 250 kbps while 802.11 operates at 2 Mbps in our experiments. Taking this into account, we normalize the Number of Nodes 90 0% BC hop delay according to the media data rate, which gives 75 25% BC 60 us a different view that the hop delay of 802.11 is 45 50% BC around 3.3 times of that of 802.15.4 (data series 802.11* 30 75% BC 15 and 802.15.4p in Fig. 6). The hop delay for sink-type 0 100% BC 1 2 3 4 application traffic is 6.3% (for 0.1 pps) to 20.9% (for 10 Attempts per Successful pps) higher than that for peer-to-peer application traffic Association (data series 802.15.4s and 802.15.4p in Fig. 6). The increment of delay is expected, since all the traffic flows now need to converge on the sink node. Fig. 9. Attempts per Successful Association vs. Beaconing Coordi- nator (BC) Ratio B. Association Efficiency TABLE I The typical scenario of an LR-WPAN is a densely S UCCESSFUL A SSOCIATION R ATE VS . B EACONING C OORDINATOR distributed unattended wireless sensor network. Self- R ATIO configuration in deployment and auto-recovery from fail- ures is a highly desirable feature in such a network [8]. Beaconing For this purpose, 802.15.4 includes an association and coordinator 0 25 50 75 100 ratio (%) disassociation mechanism together with an orphaning Successful and coordinator relocation mechanism in its design. We association 100 100 100 99 100 rate (%) give out the experimental results of association in this subsection, while the experimental results of orphaning will be given in next subsection. To associate with a coordinator, a device will perform an active channel scan, in which a beacon request frame Ratio of Devices in Beacon Mode (%) 100 100 is sent, or a passive channel scan, in which no beacon 80 67 60 52 request frame is sent, to locate a suitable coordinator. 40 22 Active channel scan is used in our experiments, since a 20 0 0 device needs to explicitly request for beacons in non- 0 25 50 75 100 beacon enabled environment. When a coordinator re- Beaconing Coordinator Ratio (%) ceives the beacon request frame, it handles it differently depending on whether itself is in beacon enabled mode or non-beacon enabled mode. If the coordinator is in Fig. 7. Devices Associated with Beaconing Coordinators beacon enabled mode, it discards the frame silently, since beacons will be bent periodically anyway. Otherwise, the coordinator needs to unicast a beacon to the device soliciting beacons. In our experiments, we vary the 1.8 percentage of beaconing coordinators to see the different effects of beaconing coordinators and non-beaconing Attempts per Association Successful 1.6 1.4 coordinators. 1.2 In general, the successful association rate is very 1.0 high (more than 99%) for different combinations of 0 25 50 75 100 beaconing coordinators and non-beaconing coordinators, Beaconing Coordinator Ratio (%) as illustrated in Table I. From Fig. 7, we can see that a device gets an almost equal chance to associate with a beaconing coordinator or a non-beaconing coordinator. Fig. 8. Association Efficiency vs. Beaconing Coordinator Ratio However, this result is obtained for beacon order 3 and it may be different for other beacon orders. Normally, a 10 TABLE II D ISTRIBUTION OF A SSOCIATION ATTEMPTS ( EXPRESSED IN NUMBER OF DEVICES ) 1 attempt 2 attempts 3 attempts 4 attempts 0% beaconing coordinators 54 30 14 2 25% beaconing coordinators 71 16 13 – 50% beaconing coordinators 79 15 5 1 75% beaconing coordinators 85 11 3 – 100% beaconing coordinators 87 11 2 – beaconing coordinator with a larger beacon order (i.e., hidden terminal problems as a fact of lacking RTS/CTS, longer superframe) reacts slowly to a beacon request, that is, even the first step of the association may fail. which means it will not get the same chance to serve The situation is better if there are multiple beaconing as a coordinator for a certain device, when competing coordinators around, since they will continue beacon- with other non-beaconing coordinators or beaconing ing as usual even if a beacon request is received. Of coordinators with smaller beacon orders. course, if beacons are sent with high frequency (low The association efficiency shown in Fig. 8, in terms beacon order), then the collisions will increase, which of attempts per successful association, is high. The will bring down the association efficiency. In summary, association procedure is a multi-step procedure as briefly non-beaconing coordinators are likely to affect the first described by the following pseudo code (for device part step of the association procedure, while the beaconing only): coordinators can affect all the steps. As revealed by our experimental results, beaconing coordinator as a whole is 1: channel scan a better choice regarding association efficiency, provided 2: if coordinators not found the beacon order is not too small. 3: association fail Table II gives out the distribution of association at- 4: elseif no coordinators permit association tempts, which shows that most of the devices succeed in 5: association fail their first association attempt, a small part of the devices 6: else try twice or three times, and three devices try four times. 7: select a proper coordinator Association is the basis of tree formation in a peer-to- 8: send association request to the coord. peer multi-hop network. The efficiency of tree formation 9: wait for ACK is directly related to association efficiency. Tree is a 10: if ACK not received useful structure and can be used by network layer, 11: association fail especially for routing purpose. In this set of experiments, 12: else a tree is quickly formed thanks to the high association 13: send data request to the coord. efficiency. Various configurations are also done during 14: wait for ACK this procedure, such as select a channel and an identifier 15: if ACK not received (ID) for the PAN, determine whether beacon enabled 16: association fail mode or non-beacon enabled mode to be used, choose 17: else the beacon order and superframe order in beacon enabled 18: wait for association response mode, assign a 16-bit short address for a device, set 19: if asso. response not received the BatteryLifeExtension option and many other options 20: association fail in the MAC layer PAN information base (MPIB). The 21: elseif association not granted smooth procedure of association and tree formation 22: association fail indicates that an 802.15.4 network has a feature of self- 23: else configuration and can shape up efficiently. 24: association succeed If there are multiple non-beaconing coordinators around, C. Orphaning they all will try to unicast a beacon, using unslotted The orphaning study is conducted in an environment CSMA-CA, to the device asking for beacons. These with all coordinators beaconing. Specifically we exam- beacons are likely to collide at the device due to the ine the orphaning behavior for different beacon orders. 11 TABLE III S UCCESSFUL A SSOCIATION R ATE VS . B EACON ORDER Beacon order 0 1 2 3 4 5 6 10 Successful association rate (%) 99 96 95 100 99 100 100 99 with high rate of orphaning, the chance an orphaned 14 device successfully recovers from all orphanings is very 12 Attempts per Association low (2% for beacon order 0 and 4% for beacon order Successful 10 8 1 as shown by data series “Devices Recovered”), but 6 4 the recovery rate of orphaning itself is not that bad 2 0 (from 30% to 89% as shown by data series “Orphanings 0 1 2 3 4 5 6 7 8 9 10 Recovered”). One point worth mentioning is that, a Beacon Order device failed to recover from all orphanings still benefits from the recovery mechanism, since its association with the coordinator is prolonged, though not to the end of Fig. 10. Association Attempts vs. Beacon order the session. D. Collision 100 Orphaning and Devices 80 Ratios (%) Recovery Orphaned 60 Devices 40 Recovered 10000 20 Orphanings Number of 7500 Collisions 0 Recovered 0 1 2 3 4 5 6 7 8 9 10 5000 2500 Beacon Order 0 0 1 2 3 4 5 6 7 8 Collisions 246859 1348 463 373 322 190 219 237 279 Beacon Order Fig. 11. Orphaning and Recovery Fig. 12. Collisions vs. Beacon Order Orphaning mechanism works only if a device is success- fully associated with a beaconing coordinator, and the device keeps tracking the beacons from the coordinator. Since orphaning is related to association, here we also Ratio of Collisions Terminals (%) give out the association results. Table III and Fig. 10 100 btw. Hidden 80 suggest that the performance of beacon enabled modes 60 with small beacon orders is not so good as that with large 40 20 beacon orders. For example, the attempts per successful 0 association for beacon order 0 is “outstanding” among 0 1 2 3 4 5 6 7 8 its peers. And the successful association rate for beacon Beacon Order order 1 and beacon order 2 is also slightly lower than others. Unsurprisingly, orphaning is also more serious in Fig. 13. Ratio of Collisions between Hidden Terminals those beacon enabled modes with smaller beacon orders (Fig. 11). The percentage of devices orphaned in beacon It is clearly shown in Fig. 12 that more collisions order 0 or beacon order 1 is about the same (around happen in low beacon orders than in high beacon orders. 58%), and is 29 times of that in beacon order 2. There is And the network virtually loses its control in beacon no orphaning in beacon order 3 or up. In an environment order 0, due to large number of collisions. This type of 12 Packet Delivery 100 100.0 Ratio of Repeated Ratio (%) 99.5 direct Collisions (%) 80 four times 99.0 indirect 60 three times 98.5 GTS twice 40 once 98.0 20 3 4 5 6 7 8 0 Beacon Order 0 1 2 3 4 5 6 7 8 four times 0.0 0.0 39. 48. 26. 54. 56. 67. 44. three times 0.0 0.0 30. 25. 42. 15. 6.8 12. 21. twice 9.5 56. 7.8 5.8 14. 3.3 12. 5.7 9.3 Fig. 16. Different Data Transmission Methods: Packet Delivery Ratio once 90. 43. 22. 20. 16. 27. 24. 14. 23. Beacon Order sions, we find the reason is that the suggested backoff length in 802.15.4 is too short, especially for long frames Fig. 14. Ratio of Repeated Collisions (Physical Protocol Data Unit larger than 100 bytes). This short backoff length results from the consideration of energy conservation, but a too short backoff length will cause repeated collisions and defeat the initial design 80 goal. The fact that no collisions repeated more than Time Ratios (%) Collision and 60 twice in beacon order 0 and beacon order 1 is somewhat coll. ratio 40 misleading. It is not because that the collisions can time ratio 20 be resolved within the first two backoffs, but that the enormous number of collisions make it impossible in 0 0 1 2 3 4 5 6 7 8 effect for a packet to collide with another packet more coll. ratio 75. 18. 21. 12. 6.5 1.6 1.3 1.4 0.0 time ratio 6.5 3.3 1.6 0.8 0.4 0.2 0.1 0.1 0 than twice before it reaches its retransmission threshold. Beacon Order The last metric we extract from this set of experiments is the time distribution of collisions within a superframe. In beacon enabled mode, a transaction (transmission of a frame as well as reception of an acknowledgment frame Fig. 15. Ratio of Collisions within the First Millisecond of a if required) using slotted CSMA-CA is required to be Superframe completed before the end of the contention access period (CAP). Otherwise, the transaction should be delayed until the beginning of next superframe. In such a design, “Beacon Storm” problem is alleviated in high order bea- more collisions are expected at the beginning of a super- cons. Due to the broadcast nature of wireless networks, frame, especially a short superframe (low beacon order) broadcast-based storm is not a rare phenomenon [18]. It in which more transactions are likely to be delayed until necessitates careful handling. the beginning of next frame. This is confirmed by our As expected, the majority of collisions happen be- experimental results shown in Fig. 15. For beacon order tween hidden terminals (Fig. 13), that is, between any 0, for example, about 75% of collisions happen within two devices not adjacent to each other in our experiments the first millisecond of a superframe (but one millisecond (see subsection IV-B). However, probability of collisions is only about 6.5% of a superframe of beacon order 0). between non-hidden terminals in low beacon orders is not trivial either. This means the slotted CSMA-CA can no longer work effectively if the beacon order is very E. Direct, Indirect and GTS Data Transmissions small, and the chance that two non-hidden terminals In this set of experiments, we compare three differ- jump to the channel simultaneously is significantly in- ent data transmission methods, i.e., direct, indirect and creased. guaranteed time slot (GTS) data transmissions (DIG). Unexpectedly, the ratio of repeated collisions is very The focus is latency (Fig. 17) and duty cycle (Fig. 18), high, as manifested in Fig. 14. By tracking these colli- but packet delivery ratio is also given (Fig. 16), for the 13 to be turned on for only about 1/64 of the duration of 2.0 a superframe, if no data to be exchanged. If the value Hop Delay (sec) 1.6 of BatteryLifeExtension is FALSE, the receiver of the direct 1.2 beaconing coordinator remains enabled for the entire indirect 0.8 CAP. In indirect data transmission, a device can enter GTS 0.4 a low power state, like sleeping state, if it finds there 0.0 3 4 5 6 7 8 are no pending packets by checking the beacon received direct 0.0055 0.0053 0.0053 0.0053 0.0053 0.0053 from its coordinator. indirect 0.0701 0.1326 0.2713 0.4953 0.9189 1.6912 GTS 0.0673 0.1108 0.1965 0.4000 0.7207 1.2935 As shown in Fig. 18, the duty cycle is around 2% in Beacon Order indirect data transmission, and about 1% in GTS data transmission. However, there are two slots or 12.5% of a superframe allocated for GTS data transmission in our Fig. 17. Different Data Transmission Methods: Hop Delay experiments, which means that (12.5 − 1)/12.5 = 92% of the allocated GTS slots are wasted. This result shows that GTS is too expensive for low data rate applications. The above duty cycle measurement is based on the 2.5 traffic load of one packet per second, and it shall Duty Cycle (%) 2.0 indirect vary when traffic load changes. Perfect synchronization 1.5 1.0 GTS among devices is also assumed in the measurement, 0.5 which is generally not true in practice. Some margin 0.0 should be provided for the non-perfect synchronization, 3 4 5 6 7 8 which means an increment in duty cycle. One more Beacon Order point about power conservation is that, it is acquired at the cost of delay, as clearly shown in Fig. 17. The power consumption mechanisms employed in 802.15.4 Fig. 18. Different Data Transmission Methods: Duty Cycle are based on the assumption of low data rate and should be used properly. sake of completion. Small beacon orders 0, 1 and 2 are not shown in the above figures, since, in GTS data VI. C ONCLUSIONS transmission, we only allocate one slot for each device At its heart, the new IEEE 802.15.4 standard, which and the slot is too short for holding a data frame. is designed for low rate wireless personal area networks No significant difference has been observed in the (LR-WPANs), is an enabling standard. It brings to light packet delivery ratio among the three data transmission a host of new applications as well as changes many methods. Nevertheless, the hop delay varies, which will other existing applications. It is the first standard to definitely affect the packet delivery ratio in upper layers. allow simple sensors and actuators to share a single The hop delay in direct data transmission is much shorter standardized wireless platform. than those in indirect and GTS data transmissions. To evaluate the general performance of this new One fundamental aspect of 802.15.4 is low power standard, we develop an NS2 simulator, which covers consumption, which is very desirable in a wireless all the 802.15.4 PHY and MAC primitives, and carry sensor network, as the replacement of batteries is very out five sets of experiments, that is, experiments of: cumbersome due to the large number of sensors. Most (1) comparing the performance between 802.15.4 and power-saving mechanisms in 802.15.4 are based on 802.11; (2) association and tree formation study; (3) beacon enabled mode. In direct data transmission, if orphaning and coordinator relocation investigation; (4) the BatteryLifeExtension option is set to TRUE, the examination of unslotted CSMA-CA and slotted CSMA- receiver of the beaconing coordinator is disabled after CA behaviors; and (5) comparing three different data macBattLifeExtPeriods (default value 6) backoff periods transmissions, namely, direct, indirect and guaranteed following the inter-frame space (IFS) period of the bea- time slot (GTS) data transmissions. Detailed experimen- con frame. Using default configuration, this means that tal results are presented, and analyses and discussions the transceiver of a coordinator or a device is required are given. 14 In non-beacon enabled mode and for low rate applica- [8] A. Cerpa and D. Estrin, “Adaptive self-configuring sensor net- tions (traffic load ≤ one packet per second), the packet works topologies,” In Proc. IEEE INFCOM, New York, June 2002. delivery ratio of 802.15.4 is similar to that of 802.11. [9] J. Heidemann, W. Ye, and D. Estrin, “An energy-efficient mac However, 802.15.4 shows clear advantage over 802.11 protocol for wireless sensor networks,” In Proceedings of the 21st regarding control overhead and transaction latency. The International Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM 2002), New York, NY, experimental results endorse the non RTS/CTS CSMA- June 2002. CA approach. [10] E. Shih, S. Cho, et al., “Physical layer driven protocol and Association and tree formation in 802.15.4 proceed algorithm design for energy-efficient wireless sensor networks,” Proc. MOBICOM, 2001. smoothly in both beacon enabled mode and non bea- [11] A. Y. Wang, S. Cho, C. G. Sodini, and A. P. Chandrakasan, con enabled mode, which implies 802.15.4 possesses a “Energy efficient modulation and MAC for asymmetric RF good self-configuration feature and is able to shape up microsensor systems,” IEEE Intl. Symp. Low Power Electronics and Design, 2001. efficiently without human intervention. The orphaning [12] V. Raghunathan, C. Schurgers, S. Park, and M. B. Srivastava, and coordinator relocation (recovery from orphaning) “Energy-aware wireless microsensor networks,” IEEE Signal mechanism provides for a device a chance of self-healing Processing Magazine, Volume:19 Issue: 2, March 2002 Page(s): 40-50. from disruptions. The orphaning recovery probability is [13] C. Karlof and D. Wagner, “Secure routing in wireless sensor net- about 30% for the worst case and about 89% for the best works: attacks and countermeasures,” In First IEEE International case in our experiments. Notwithstanding, the chance Workshop on Sensor Network Protocols and Applications, 2003. [14] A. Perrig, R. Canetti, D. Song, and D. Tygar, “The TESLA that an orphaned device is completely recovered, that broadcast authentication protocol,” In RSA Cryptobytes, Summer is, it recovers each time it is orphaned, is very low. 2002. For the lack of RTS/CTS, 802.15.4 is expected to [15] Y. Hu, D. B. Johnson, and A. Perrig, “SEAD: Secure efficient distance vector routing for mobile wireless ad hoc networks,” suffer from hidden terminal problems. Our experiment In Proceedings of the 4th IEEE Workshop on Mobile Comput- results match this expectation. But for low data rates up ing Systems & Applications (WMCSA 2002), pp. 3-13, IEEE, to one packet per second, the performance degradation Calicoon, NY, June 2002. [16] L. Eschenauer and V. Gligor, “A key-management scheme for is minor. The default CSMA-CA backoff period in distributed sensor networks,” Conference on Computer and Com- 802.15.4 is too short, which leads to frequent repeated munications Security. Proceedings of the 9th ACM conference on collisions. Superframes with low beacon orders can Computer and communications security, Washington, DC, 2002. [17] D. Ganesan, B. Krishnamachari, et al., “Complex behavior at also lower the slotted CSMA-CA backoff efficiency and scale: An experimental study of low-power wireless sensor lead to high collision probability at the beginnings of networks,” UCLA/CSD-TR 02-0013, UCLA Computer Science, superframes. 2002. [18] S. Ni, Y. Tseng, Y. Chen, and J. Sheu, “The broadcast storm Our study shows that 802.15.4 is an energy-efficient problem in a mobile ad hoc network,” In Proceedings of the fifth standard favoring low data rate and low power consump- annual ACM/IEEE international conference on Mobile computing tion applications. GTS data transmission is an expensive and networking, pages 151-162, ACM Press, 1999. [19] A. Woo and D. Culler, “A transmission control scheme for media approach for low data rate applications, as can be seen access in sensor networks,” In Proceedings of the Seventh Annual from our experimental results. International Conference on Mobile Computing and Networking (MOBICOM-01), pages 221-235, ACM Press, New York, 2001. [20] R. D. Pietro, L. V. Mancini, and A. Mei, “Random key assign- R EFERENCES ment for secure wireless sensor networks,” In Proceedings of the [1] IEEE P802.15.4/D18, Draft Standard: Low Rate Wireless Per- 2003 ACM Workshop on Security of Ad Hoc and Sensor Networks sonal Area Networks, Feb. 2003. (SASN ’03), October 2003. [2] IEEE 802.11, Part 11: Wireless LAN medium access control [21] M. Zorzi and R. R. Rao, “Multihop performance of energy- (MAC) and physical layer (PHY) specifications, IEEE, Aug. efficient forwarding for ad hoc and sensor networks in the pres- 1999. ence of fading,” IEEE/ICC’04, 20-24 June 2004, Paris, France. [3] Bluetooth SIG, Bluetooth Specifications, V1.0, July 1999. [22] A. D. Wood and J. A. Stankovic, “Denial of service in sensor [4] FIPS Pub 197, Advanced Encryption Standard (AES), Federal networks,” IEEE Computer Magazine, October 2002, pp. 54-62. Information Processing Standards Publication 197, US Depart- [23] J. H. Schiller, Mobile Communications, Addison-Wesley, 2000. ment of Commerce/N.I.S.T, Springfield, Virginia, November 26, 2001. (http://csrc.nist.gov/) [5] USC Information Sciences Institute, Marina del Rey, CA. Net- work Simulator – NS2. (http://www.isi.edu/nsnam/ns). [6] USC Information Sciences Institute, Marina del Rey, CA. Net- work Animator – Nam. (http://www.isi.edu/nsnam/nam). [7] Y. Liu, M. Lee, and T. Saadawi, “A bluetooth scatternet-route structure for multi-hop ad hoc networks,” IEEE Journal on Select Areas in Communications, Vol. 21, No. 2, pp.229-239, Feb. 2003.