Performance of 802.11 DCF in Clustered Ad Hoc Sensor

Document Sample
Performance of 802.11 DCF in Clustered Ad Hoc Sensor Powered By Docstoc
					                              Performance of 802.11 DCF in Clustered
                                     Ad Hoc Sensor Networks

                       Qingjiang Tian, Seema Bandyopadhyay and Edward J. Coyle
                              School of Electrical and Computer Engineering
                           Purdue University, West Lafayette, IN 47907-2035
                                 {tianq, seema, coyle}@ecn.purdue.edu


                      Abstract                                    The delay in collecting data from all sensors in a
                                                            sensor network is another important issue. For
 In this paper, we study the performance of a modified      example, in target tracking applications, immediate
802.11 DCF under a multiple-hop cluster scenario in ad      detection and rapid response to an event is critical.
hoc sensor networks. We provide simulation results          Delays in gathering target data from sensors to a
regarding the time to gather one packet from each           processing center must thus be minimized.
sensor in a cluster. The trade-off between this gathering         802.11 is the dominant MAC protocol in current
time and the energy expended by the sensors during this     wireless networks. In this protocol, two fundamental
time period is studied. Comparisons between 802.11          mechanisms to access the medium are defined. The
DCF and an idealized transmission scheduling protocol       most common one is called the Distributed
are provided. We also address the issue of non-uniform      Coordination Function (DCF), a random access
spatial sampling resulting from the use of 802.11 DCF       scheme that is based on carrier sense multiple access
in multiple-hop sensor clusters.                            with collision avoidance (CSMA/CA). The alternative
                                                            mechanism is called the Point Coordination Function
                                                            (PCF), a centralized MAC protocol that is able to
Keywords: Ad Hoc Sensor Networks, 802.11 DCF.               support collision-free and time -bounded services.
                                                                 While 802.11 has been proposed mainly for
                 I. Introduction                            homogeneous peer wireless nodes communication,
                                                            little work has been done on investigating the
     Recent developments in the technologies of             performance of 802.11 in clustered ad hoc sensor
wireless communications, computing and micro-               networks. In this paper, we thoroughly study the
electro-mechanical systems have enabled the creation        performance of a modified 802.11 DCF in an ad hoc
of extremely small, low-cost sensors. These sensors         sensor network scenario. The main contributions of
can be deployed in large numbers and organized into         this paper are as follows:
ad-hoc networks that can perform such tasks as                 • The trade-off between time delay and energy
environmental monitoring, target tracking, and system               usage when collecting one data packet from
control.                                                            each node in a network.
     The small size of these sensors limits the energy         • Demonstration of the non-uniform spatial
that is available to them, both from batteries and on-              sampling that results when 802.11 is used in
board devices that gather energy from the environment.              DCF mode.
This limits the power available for communications,            • Comparisons between 802.11 and a scheduling
thus     constraining   their   transmission     range,             protocol for which analytical results are
transmission duration and bit rate. Energy-efficiency               available [7].
has thus been a key consideration in research and                The rest of the paper is organized as follows:
development efforts for these sensor networks. Key          Related work is discussed in Section 2. In Section 3,
contributions in this area have included scheduled          we summarize the system model and our assumptions.
sleep times to conserve energy [11], clustering             The 802.11 configuration used in all simulations in
algorithms that minimize the number of transmissions        this paper is given in Section 4. Simulation results and
[5], and protocols that minimize collisions of              their comparisons with known analytical results are
transmissions [6].
presented in Section 5. Conclusions are provided in          environment. The sensors in this network gather data
Section 6.                                                   related to the monitored environment and
                                                             communicate it to a processing center where it is used
                  II. Background                             to estimate parameters characterizing the monitored
                                                             environment. The processing center is a specialized
                                                             device or one of the sensors itself. The sensors in the
     Since 802.11 was proposed in 1997, its                  network are organized into clusters as proposed in [5].
performance has been analyzed in a number of                     In each cluster, one of the sensors acts as a CH and
different scenarios [8][9]. In [8], the approximation of     is responsible for collecting data from all other sensors
a constant and independent collision probability             in its cluster, aggregating it and then forwarding to the
suffered by each transmitted packet is used in the           processing center. Since the sensors communicate data
construction of a Markov chain model of the protocol.        to the processing center through the CHs, the data has
This model is then used to numerically analyze the           to travel smaller distances, minimizing the total energy
saturation throughput of 802.11 DCF. In [9], an              spent in gathering data form network.
M/M/1/k model of 802.11 DCF with Poisson arrival                  We make following assumptions regarding the
traffic is developed and analyzed. In this model, the        sensor network.
saturation throughput results in [8] are used to justify        a) The sensors locations form a homogeneous
approximating the departure process as a Poisson                     spatial Poisson process of intensity λ in a
process. While these efforts have produced fairly
                                                                     square of dimension 2a × 2a .
accurate results on the performance of 802.11 DCF,
                                                                b) All sensors have the same transmission range:
they are all based on the assumption of homogeneous
                                                                      r . Communications between two nodes further
peer-to-peer communications among wireless nodes
                                                                     than r from each other will be forwarded by
within one hop of each other.
                                                                     other nodes.
     Different approaches have been proposed to
                                                                c) Two nodes separated by a distance larger than
minimize the energy spent in gathering data from
                                                                      2r will not interfere with each other.
nodes in ad hoc sensor networks. Among them,
                                                                d) Each packet contains one data sample. The time
clustering appears to be a very promising approach [5].
                                                                     to gather one packet from each sensor is thus
In [5], a multiple -hop hierarchical clustering algorithm
                                                                     the inverse of the networks temporal sampling
has been proposed. In this approach, each node will
have a certain probability of volunteering to become a               rate.
cluster-head (CH). The maximum number of hops in
each cluster is bounded and those nodes that do not                  IV. Simulation Configuration
join some volunteer CH’s cluster are then forced to
become isolated clusterheads. Proper choice of the            4.1. Summary of 802.11 DCF
maximum number of hops in a cluster will minimize
the number of forced CHs. Under this algorithm,
                                                                     The IEEE 802.11 standard defines two modes of
multiple-hop clusters will cover the monitored area.
                                                               operations: Distributed Coordinated Function (DCF)
Analytical results show that this clustering algorithm
                                                               and Point Coordinated Function (PCF). The latter
is much more energy-efficient than non-clustered or
                                                               mode is a centralized MAC protocol that supports
single-hop cluster algorithms.
                                                               collision-free and time-bounded services. In this paper,
    A scheduling protocol has been proposed for
                                                               we focus on 802.11’s DCF mode.
collecting data from nodes in clustered ad hoc sensor
                                                                      In DCF mode, a node that has just generated a
network [6]. Based on the physical location
                                                               data packet to transmit first senses the channel for
information of non-CH nodes within the cluster, the
                                                               activity. If the channel is idle for a period of time
CH        schedules      simultaneous       collision-free
                                                               called the distributed interframe space (DIFS), the
transmissions. This work also provides a lower bound
                                                               node will begin transmitting. If the channel is sensed
on the time to empty the cluster – the time to gather
                                                               to be busy at any time during the DIFS period, the
one packet from each sensor in the cluster.
                                                               node awaits the return of the channel to the idle state
   The results in [6] are useful when e     valuating the
                                                               and then restarts the DIFS countdown. At the end of a
merits of other protocols. They will provide the
                                                               DIFS after a busy channel, the node generates a
baseline against which we compare 802.11 DCF.
                                                               random backoff time interval and waits until the end
                                                               of this interval before attempting a transmission. This
  III. Assumptions & Network Topology                          randomized backoff is designed to reduce the
   We consider a wireless network of extremely small           probability of a collision between two or more nodes
and inexpensive sensors deployed to monitor a certain          that are each awaiting the end of a busy channel
interval. To avoid channel hogging, a station is also       handshake. We will only consider this four–way
required to generate a random backoff interval              handshaking mechanism and will determine its energy
between any two consecutive new packets’                    overhead as the packet length is varied.
transmissions, even when the channel is sensed idle
for one DIFS.                                               4.2. Modifications to NS
      The 802.11 DCF backoff time interval is slotted.
The slot size, which depends on both the physical                All of our simulations are conducted with NS-2
layer and MAC layer characteristics, is set equal to the    [13]. In this subsection, we briefly discuss our
time needed at any node to detect the transmission of a     modification to the current NS-2 implementation of
packet from any other node. DCF adopts an                   802.11 DCF.
exponential backoff scheme. At each packet                        The purpose of our simulations is to study the
transmission, the backoff time is uniformly chosen in       idealized performance of 802.11 DCF in a multi-hop
the range of (0, w -1) slots, where w is called the         cluster scenario in the ad hoc sensor network proposed
contention window and depends on the number of              in [5]. For reasons of simplicity, only one cluster is
transmission failures for the packet. Each failure will     considered in all simulations; however, it is
result in the contention window being doubled, up to        straightforward to extend our simulations to the case
 CWmax = 2 m CWmin , where m is the backoff stage.          of multiple clusters. In our simulations, we consider a
      The backoff counter for a packet is decremented       single K -hop, circular cluster with node density λ .
once for each slot that the channel is idle and freezes     The CH is assumed to be located at the center of the
when the channel is busy. The countdown resumes             cluster.
when the channel is sensed idle again for a DIFS                  The underlying network routing protocol is the
period. The node starts to transmit the packet when the     Destination-Sequenced        Distance-Vector      Routing
backoff time counter reaches zero.                          (DSDV) protocol [12]. DSDV is a table -driven
       The standard also defines two different access       algorithm based on the classical Bellman-Ford routing
modes for DCF. In one mode, one node sends out a            mechanism. Each node will maintain a routing table
data packet and the receiving node sends back an ACK        that contains all of the possible destinations within the
after a period of time called the short interframe space    network. Each entry in the routing table is marked
(SIFS). This two-way handshaking technique for              with a sequence number assigned by the destination
successful data packet transmissions is called the basic    node, which enables the mobile nodes to distinguish
access mechanism.                                           stale routes from new ones, thereby avoiding the
    In the second mode, the transmitting node will first    formation of routing loops. Routing table updates are
send out a control packet called a Request To Send          periodically transmitted throughout the network to
(RTS). When the receiving node detects this RTS             maintain table consistency. Two types of update
frame, it responds with a Clear To Send (CTS) packet.       packets are possible, the first one is known as a full
When the transmitting node receives the CTS, it             dump, which carries all available routing information.
follows the operations defined in the basic access          Smaller incremental packets are used to relay only that
mechanism. The advantage of this four-way                   information that has changed since the last full dump.
handshake -- the RTS-CTS handshake followed by the          While the second type of update packet must fit into a
basic access handshake -- is that it ensures the receiver   standard-size Network Protocol Data Unit (NPDU),
is ready and reserves the channel for the packet            the full dump update packet may be of multiple
transmission.                                               NPDUs.
      Both the RTS and CTS contain the length of the              Our simulation is divided into two steps. Step 1,
data packet to be transmitted, and all nodes within         called the training step, lasts for T1 units of simulation
range of both the transmitter and receiver read this        time. T1 is chosen large enough that nodes’
information. Each of these nodes uses this information      transmissions can be spread out to minimize the
to set their network allocation vector (NAV), which         probability of collisions. During the period [0 T ],    1
indicates the period of time during which the channel       each node will send one packet to the CH. The
will remain busy – and during which they will remain        purpose of this first step is to enable each node in the
idle.                                                       simulator to determine, via the mechanism described
     The RTS-CTS mechanism is very effective in             above, a routing path to the CH.
improving system performance in multi-hop networks,              At the beginning of this training step, none of the
especially when packet lengths are significantly longer     nodes has knowledge of its neighbors, so each node
than the DIFS. The reason is that helps mitigate the        will broadcast an ARP (Address Resolution Protocol)
hidden-terminal problem via the reservation                 query packet when it has a data packet to transmit. In
                                                            the current NS-2 implementation, the ARP module of
each node can only cache one data packet for which                 Before presenting our simulation results, all
the destination address is unavailable. To guarantee           variables in the simulation are summarized as follows:
that each node within the cluster will acquire the path             K : Maximum number of hops from a sensor to
to the CH, we modified this buffer size to                         the CH
accommodate all possible waiting data packets for the               Ts : Random Start Time
ARP module in arp.h. We also modified arp.cc to                     N : Total number of nodes within the cluster
guarantee that each waiting data packet will be                     T E : Time to empty the cluster once
transmitted properly when the destination node
                                                                    Eo : Transmission Energy Overhead
address is available.
                                                                    λ : Node Density
       Step 2 of the simulation is used to study the
idealized performance of 802.11 DCF under a                      The simulation parameters of 802.11 DCF in NS-
                                                             2 were set as shown in Table 1.
clustered scenario. As mentioned above, DSDV will
generate periodic route update packets. To eliminate
                                                                          Table 1. Simulation configuration
the effect of such “control overhead” from our delay
                                                                       CW Min                       32
and energy computations, we modified the
implementation of DSDV in NS-2. Since (a) each                         CW Max                     1024
node has already acquired the path to the CH in Step 1,                  SIFS                      10us
(b) the CH is the only destination for all the non-CH                 Slot time                    10us
nodes in the cluster, and (c) the location of each node               Data Rate                  1Mbps
is fixed throughout the simulation, we modified                          RTS                    352(bits)
dsdv.cc such that neither type of update packet -- full
                                                                         CTS                    304(bits)
dump packets or smaller incremental packets -- will be
sent after T1 .                                                          ACK                    304(bits)
                                                                  Packet Length (PL)         1000, 3000(bits)
           V. Simulation Results
                                                                    Fig.1 and Fig.2 show the performance comparison
                                                               between 802.11 DCF and the collision-free protocol
     In this section, we provide our simulation results
                                                               analyzed in [6]. The node density in Fig. 1 is 11; in
for 802.11 DCF. Comparisons between 802.11 DCF
                                                               Fig. 2 it is 5.6. In both figures, the maximum number
and the bound derived in [6] are provided. We also
                                                               of hops was set to K =3. The data packet length was
show the trade-off between energy overhead and time
                                                               set to PL=3000bits, which is equivalent to 0.003us at a
delay for 802.11 DCF. The spatial density of sensors
                                                               data rate of 1Mbps. In both figures, time is measured
from which packets are actually received at the
                                                               in multiples of the data packet length.
clusterhead is also determined.
                                                                      We simulated and traced the time for the CH to
    The simulation configuration is as follows: N
                                                               collect packets from a certain percentage of non-CH
nodes are located within a circle of A = πk 2 square           nodes. The lower bound in each figure shows the time
units (here we assume the transmission range is 1).            required by the collision-free algorithm in [6] to
Thus, the cluster is a complete K -hop cluster and the         completely empty the cluster of packets (100%
CH is located at the center of the cluster.                    collection). From Fig.1 and Fig.2, it is clear that the
     The CH will collect exactly one packet from each          delay for the collision-free algorithm is approximately
node within the cluster. For each i , the i’th node starts     1/2 the delay of 802.11 DCF. The reason is obvious:
its transmission at t i . The ti ’s are independent and        the random access to the common channel in 802.11
uniformly distributed in the interval [0 Ts ]. Ts ,            DCF will unavoidably result in RTS collisions and
which is called the random start time, plays an                random backoffs, both of which increase the time
important role in the performance of 802.11 DCF. If            required to collect packets.
 Ts is zero, the most extreme case, all the nodes in the             In Fig.1, it is obvious that the optimal choice for
                                                                Ts is 1000 time units (a unit is the time to transmit
cluster will sense the channel idle for DIFS, then start
their transmissions. Obviously, all the packets will           one packet). As mentioned above, any choice of Ts
collide and be lost. On the other hand, if Ts is too           represents a trade-off between the number of RTS
large, there will be very long delays between                  collisions and the idle time between successive
successive packets transmissions. Hence, there exists a        transmissions. For the optimal random start time
finite value of Ts that minimizes the time to empty the        shown in Figure 1, the expected number of RTS
cluster.                                                       collisions per successfully transmitted packet was 3.
                                                             Fig. 3: Transmission energy overhead as a function of
   Fig. 1: Delays in gathering packets from specified                the random start time. λ =5.6 and K=3
fractions of all sensors when 802.11 DCF is used. The
 lower bound is the delay in gathering packets for the
     collision-free protocol in [6]. λ =11, K=3 and
                      PL=3000(bits).




                                                               Fig.4: Trade off between transmission overhead and
                                                             time to collect packets from a cluster. λ =5.6 and K=3

                                                                  We assume that the sensors are uniformly
Fig.2: The same situation as in Figure 1, except λ =5.6      distributed over the target region. If one data sample
                                                             can be collected from each sensor, then the samples
    In ad hoc sensor networks, energy consumption is         are also uniformly distributed and have the same
a very important issue. In Fig. 3, we plot the               density over the region. The use of clustering in a
transmission energy overhead for two different packet        sensor network will result in uneven sampling because
lengths. Here, we assume the energy to transmit a data       the samples from sensors near the clusterhead are
packet is “useful” while the energy for RTS/CTS/ACK          collected more easily than those that are multiple hops
packets is overhead. While the two data packet lengths       away from the clusterhead.
in the figure are both less than the fragmentation                  Figs. 5 and 6 show this non-uniform sampling
threshold, it is clear that the transmission overhead        phenomenon for a cluster with K =4 hops. The
becomes worse when the data packet becomes smalle r.         simulation configuration in both cases is similar to the
Furthermore, when the random start time Ts becomes           previous ones except that K =4 instead of K =3.
large, the transmission overhead decreases because                In Fig.5, we determine sampling densities in
there are fewer collisions. When Ts is so large that         different rings of the cluster when the sensor densities
collisions are extremely rare, the overhead is simply        are λ =5 and 10. The lowest set of curves in the figure
one RTS/CTS/ACK sequence for each successful
                                                             show that when the CH collects 10 percent of all
packet.
                                                             packets for the λ =5 case, 53% packets in the first hop
     Fig. 4 shows the trade-off between the
                                                             have been sampled while less than 20 percent packets
transmission energy overhead and the time to collect
                                                             are sampled in the second hop. When λ =10, around
packets from a certain percentage of sensors in the
cluster. It clearly shows that reducing the transmission     65% nodes in the first hop have been sampled once,
overhead by increasing the random start time Ts will         but less than 4% of nodes in the fourth hop.
increase the time to collect packets from the cluster [7].
     We can see, (a) the CH favors nodes closer to it                      VI. Conclusions
under 802.11 DCF and (b) higher node densities
accentuate this trend.                                         This paper described the performance of 802.11
                                                         DCF in clustered ad hoc sensor networks. Simulation
                 Upper bound                             was used to determine: (a) the trade-off between the
                     90 percent                          time to gather packets and energy overhead expended
                                                         in gathering them; (b) the effect of the random start
                                                         time on performance of the cluster; and (c) the non-
                                                         uniform sampling caused by clustering. Where
                                            70 percent
                                                         applicable, these results were compared with those for
                 50 percent                              a collision-free protocol, thus allowing insight into the
                                                         performance compromises made when 802.11 DCF is
                                                         used.
                               10 percent
                                                         References
                                                         [1]    http://www.sac.eecs.berkeley.edu/~warneke/pubs/index
Fig. 5: 802.11 Sampling density distribution with node          .html
              density λ =5 and 10, K=4                   [2]    http://bwrc.eecs.berkeley.edu/Research/Pico_radio/Def
                                                                ault.html.
      In Fig. 6, we compare the percentages of nodes     [3]    http://cens.ucla.edu/.
from which packets have been collected after a fixed
                                                         [4]    I. F. Akyildiz, W. Su, Y. Sankarasubramaniam and E.
time using 802.11 DCF versus the scheduling                     Cayirci, "Wireless Sensor Networks: A Survey",
algorithm [6, 7]. It is apparent from Fig. 6 that the           Computer Networks (Elsevier) Journal, pp.393-422,
scheduling algorithm is more time-efficient than                March 2002.
802.11 DCF. At time T1 , when the CH has collected       [5]    S. Bandyopadhyay and E. J. Coyle, “An Energy
packets from 90% of the nodes that are one hop away             Efficient Hierarchical Clustering Algorithm for
using the scheduling algorithm, it has collected                Wireless Sensor Networks”, In Proc. INFOCOM’03,
packets from only 60% of the nodes that are one hop             San Francisco, April 2003.
away using 802.11 DCF. Moreover, we also observe         [6]    S. Bandyopadhyay and E. J. Coyle, “Spatio-Temporal
that the sampling non-uniformity is more pronounced             Sampling Rates and Energy Efficiency in Wireless
when 802.11 DCF is used versus the scheduling                   Sensor Networks,” Submitted to INFOCOM’04.
algorithm. The random access mechanism associated        [7]    S. Bandyopadhyay, Q. J. Tian and E. J. Coyle, “Spatio-
with 802.11 DCF results in unavoidable transmission             mporal Sampling Rates for Wireless Sensor Networks”.
control overhead and collisions, which leads to the             Accepted by Fall VTC’2003.
time inefficiency compared to the scheduling             [8]    G. Bianchi, “Performance Analysis of the IEEE 802.11
algorithm.                                                      Distributed Coordination Function”, in J. Sel. Area.
                                                                Comm, Vol. 18, No. 3, pp.535-547.
                                                         [9]    H. Foh and M. Zukerman, “Performance evaluation of
                                                                IEEE 802.11”, In Proc. of EW2002 Vol.2, pp.841-845.
                                                         [10]   IEEE, “IEEE Standard for Information Technology –
                                                                Telecommunications and Information Exchange
                                                                between Systems – Specific Requirements – Part 11:
                                                                Wireless LAN MAC and PHY Specifications,” IEEE
                                                                Std 802.11-1999, IEEE, New York, 1999.
                                                         [11]   W. Ye, J. Heidemann and D. Estrin, “An Energy -
                                                                Efficient MAC Protocol for Wireless Sensor
                                                                Networks”, In Proc. INFOCOM’02, New York, June,
                                                                2002.
                                                         [12]   C. E. Perkins and P. Bhagwat, “Highly Dynamic
                                                                Destination-Sequenced       Distance-Vector    Routing
                                                                (DSDV) for Mobile Computers”, Comp. Commun.
  Fig. 6: Sampling density distribution comparison              Rev., Oct., 1994, pp.234-44.
between 802.11 DCF and the scheduling algorithm at       [13]   http://www-nrg.ee.lbl.gov/ns/#version2.
                    fixed time T