Download A Context-Aware Cross-Layer Broadcast Model

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Network tap wikipedia , lookup

Net bias wikipedia , lookup

Wireless security wikipedia , lookup

Piggybacking (Internet access) wikipedia , lookup

Computer network wikipedia , lookup

Backpressure routing wikipedia , lookup

IEEE 1355 wikipedia , lookup

Deep packet inspection wikipedia , lookup

Wake-on-LAN wikipedia , lookup

Recursive InterNetwork Architecture (RINA) wikipedia , lookup

IEEE 802.1aq wikipedia , lookup

CAN bus wikipedia , lookup

Airborne Networking wikipedia , lookup

Cracking of wireless networks wikipedia , lookup

Peer-to-peer wikipedia , lookup

Routing wikipedia , lookup

Kademlia wikipedia , lookup

Routing in delay-tolerant networking wikipedia , lookup

Transcript
A Context-Aware Cross-Layer Broadcast Model for Ad
Hoc Networks
Ouksel, Aris and Lundquist, Doug (corresponding author)
University of Illinois at Chicago, 601 S. Morgan St., Chicago, Illinois, 60607, USA
[email protected], [email protected]
Abstract—The standard 802.11 medium access control (MAC) performs poorly for heavy broadcast traffic. We
present our context-aware cross-layer (CACL) broadcast model as an alternative. The basic CACL model uses only
contextual data available to the 802.11 MAC and so is usable by any routing protocol that uses the 802.11 MAC.
CACL fits the total broadcasts in any two-hop neighborhood to wireless channel capacity. We compare collision
rates for CACL and the 802.11 MAC and conclude that, for a wide range of network conditions, CACL offers
superior single-hop transmission rates. We also present a geographically-constrained extension to CACL, CACL-G,
and compare it against CACL in vehicular scenarios of varying node density. Our results show that CACL-G offers
increasingly superior performance over the basic CACL model as node density increases.
Keywords—Ad hoc networking; broadcasting; vehicular networks; multi-hop routing; simulation
1 Introduction
Broadcasting is a core communication method for ad hoc networks and is useful under a wide range of network
conditions. Ad hoc networks may be composed of static sensors, low-mobility pedestrians, high-mobility vehicles,
or combinations thereof. Communication may be primarily one-to-one, one-to-many, many-to-one, or many-tomany. In all combinations of mobility and communication models, broadcasting is used to learn routes, deliver
information to unknown destinations, and avoid redundant transmissions of popular content, as in [32]. Broadcasting
assumes even more importance for sharing information among large groups of mobile nodes because building routes
between specific node pairs may be impractical.
Routing protocols such as Dynamic Source Routing (DSR) and Ad Hoc On-Demand Distance Vector Routing
(AODV) build routes by flooding route request packets in the area around a source. In static topologies,
broadcasting may only be required to establish routes, which can be reused thereafter. However, when nodes quickly
move out of mutual communication range, any learned routes are ephemeral. Mobility’s negative effects on routing
path persistence are exacerbated by heavier network traffic, which increases forwarding delays at each hop on a
routing path. With increasing mobility and network traffic, routes must be built at shorter intervals. Ultimately,
flooding via broadcast becomes the only way to deliver packets quickly and reliably.
Although reliable, flooding is intrinsically bandwidth-intensive, more so as network size and activity increase.
Unregulated (naïve) flooding, where every node forwards every new packet it receives, leads to the well-known
broadcast storm problem. Broadcast storm wastes bandwidth by high collision rates and excessively redundant
transmissions, where many nearby nodes retransmit the same data to the same recipients. The three basic solutions
to this problem are limiting the receipt of new packets (by regulating flooding depth), limiting the transmission rate
(by increasing the delay between broadcasts), and by limiting saturation (where not all nodes forward a given
packet). Any large-scale broadcast-reliant application must apply at least one of these flooding controls.
Fast and efficient controlled flooding allows more current knowledge and more frequent updates for group learning
applications. Proposed applications involving direct data exchanges between vehicles include pollution credit
trading [31], vehicular safety [34], and position estimate correction [11]. The crucial commonality in the preceding
three examples is that they share information within geo-social vehicle groups (i.e., all participating vehicles in a
limited area). For such applications, information will be exchanged primarily via broadcast-reliant controlled
flooding. Efficient bandwidth usage is a vital concern for both short-range wireless standards like 802.11 a/b/g/n
standards and the longer-range DSRC (using 802.11p). These communication models all include the 802.11 medium
access control (MAC), extended for 802.11p to enable fast interactions typical of vehicular applications.
Unfortunately, the 802.11 MAC performs poorly under heavy broadcast traffic [28].
This well-known limitation of the 802.11 MAC arises because contention for the wireless channel is handled the
same way regardless of network conditions. In applications where nodes only broadcast occasionally, the problem is
minor but it becomes very important when broadcasting is the sole or dominant communication method. For such
broadcast-reliant applications, network performance can be substantially improved by context-awareness, where
nodes learn about local conditions and adapt their behavior accordingly. For example, high node density tends to
induce excessive redundancy in flooding. Context-awareness might address this issue by having only a fraction of
nodes in a high-density area broadcast a given packet.
We address the broad research question of how context-awareness can be used to improve network performance.
Group learning applications inherently gather local contextual information, which can be applied to reduce collision
rates and improve end-to-end packet delivery ratios and routing time. Our ongoing research has proposed methods
for using contextual knowledge to improve network performance. This paper extends our earlier work in contextaware broadcast scheduling and routing protocols. Our first set of simulation results is for single-hop collision rates,
varying node density and network traffic. Our second set of simulation results is for multi-hop routing in vehicular
environments, varying node density.
Single-hop simulations: We compare our Context-Aware Cross-Layer (CACL) broadcast scheduling scheme
against the standard 802.11 medium access control [7]. CACL uses each node’s knowledge of its outgoing
packets and one-hop neighbors to regulate its broadcast rate and probabilistically avoid collisions. This context
is available to the 802.11 MAC from normal packet handling and so CACL can be used by any routing protocol
that uses the 802.11 MAC for broadcasts. For any node n, CACL controls n’s broadcast rate in accordance with
its estimated two-hop neighborhood population, giving a stable collision rate. In contrast, the 802.11 MAC
model accepts packets from the routing layer without regard to either node density or network activity, which
can lead to high collision rates. These results were presented previously in [30].
Multi-hop simulations: Extending our standard CACL model, we compare it against CACL-G, which limits
flooding areas according to geographical constraints. CACL-G only forwards those packets within a polygon
defined by each packet’s source and destination. For this comparison, we use the medium access controls in
conjunction with the Self-Balancing Supply/Demand (SBSD) routing protocol [30]. SBSD controls flooding
using a utility metric that includes time, distance, and popularity and provides for redundant transmissions for
delay tolerance. We compare the results obtained by varying the geographical flooding constraints of CACL-G
in addition to the constraints already enforced by SBSD.
The rest of this paper is structured as follows. Section 2 discusses the limitations of the 802.11 MAC, and presents
our CACL model. Simulation results for single-hop collision rates are presented and discussed in section 3 and for
multi-hop routing in section 4. Section 5 discusses additional relevant research. Finally, we offer a conclusion and a
few directions for future research in section 6.
2 Cross-Layer Schemes
The IEEE 802.11 MAC performs quite well in many wireless network settings. When network traffic is low,
throughput under the 802.11 MAC approaches the theoretical ideal [6]. Further, its collision-avoidant policy
generally improves power-efficiency; this is important when mobile devices rely on small batteries. However,
problems arise under heavy network traffic, since each node transmits packets without regard to whether receiving
nodes can forward them. Severe mismatches between network traffic and channel capacity are an important
scalability problem in large networks [3] and even in small networks with high message posting rates.
Heavy network traffic is a natural consequence of high node density for group learning applications, as more nodes
generate and transmit data. It was shown in [10] that broadcast collision rates for the 802.11 MAC increase with
rising node density, further exacerbating network congestion. Under persistent heavy traffic, nodes using the 802.11
MAC may drop many incoming packets, with negative effects on end-to-end delivery. Hidden terminal collisions
can become common and sharply reduce throughput [16]. As the network becomes congested, nodes accumulate a
backlog of packets awaiting transmission; any node that claims the channel may continue transmitting for relatively
long periods. This is especially harmful for non-uniform densities, as the 802.11 MAC does not guarantee that nodes
in low-density areas between high-density areas will regularly be able to transmit.
Cross-layer approaches can address the above limitations of the 802.11 MAC by adapting node behaviors to local
conditions. Awareness of density enables nodes to limit their broadcast rate, avoiding excessive network congestion.
It also can enforce forwarding diversity, by ensuring all nodes endure a density-based delay for transmitting again;
nodes in sparse areas will broadcast more often. For example, density estimates were used for selective forwarding
[22], [23] and to regulate exponential delays between broadcasts [24]. As we will show in section 2.2, our CACL
model is designed so that its broadcast collision rate converges to 1/e with increasing node density. This bounded
collision rate enables superior network throughput compared to the 802.11 MAC for broadcast-reliant applications
with heavy network traffic.
2.1 Context-Aware Cross-Layer Model
CACL resides (Fig. 1) at the MAC layer, where it maintains knowledge of node density that may be shared with the
routing layer. The routing layer passes packets to CACL for broadcast (although Fig. 1 shows the queue of packets
residing at the routing layer, this residence is not essential for CACL). After each completed broadcast, CACL calls
the routing layer, requesting the next packet to be broadcast and informing the routing layer when that broadcast will
occur. From handling packets, CACL at a node n learns three contextual variables: node density around n; the
duration of the previous transmission by n; and n’s current state (i.e., idle, transmitting, or receiving).
Fig. 1. Information storage and flows.
After a node n broadcasts, CACL imposes a delay before n can broadcast again. The delay is based on the estimated
two-hop node population around n since collisions can occur if nodes within two hops of each other transmit
simultaneously. After a broadcast requiring time t by a node n with estimated one-hop neighborhood N1 and two-hop
neighborhood N2, CACL imposes an average delay of 2(N2 – N1)t before n may broadcast again, randomly taken
from the range [(N2 – N1)t, 3(N2 – N1)t]. Like the 802.11 MAC, CACL prevents nodes from transmitting while they
are receiving a transmission. This prevents collisions between nodes in mutual communication range. If n is
receiving a transmission when it is scheduled to broadcast, it checks at intervals averaging 0.1ms that it is idle before
broadcasting.
The one-hop neighborhood N1 is independently learned by each node n by reading the sender’s address in each
packet’s MAC header. After each broadcast by n, N1 is checked for accuracy; if n has not received a packet from a
node in its N1 list within the last 500ms, that node is purged from the list. We chose this cutoff time because, given
our simulation parameters, nodes are expected to broadcast several times during it. It therefore provides an accurate
estimate of N1. Though not explored in this paper, in principle cutoff times could be adaptively set to allow any
desired number of expected transmissions given a known density and packet size distribution.
Each node estimates its two-hop neighborhood N2 linearly from its knowledge of N1. In [29], node mobility was not
constrained by roads and so N2 was estimated as 4N1. That is, N2 occupied a circular area of twice the radius (and
four times the area) as N1. Under road-based mobility, the node population along a road segment is estimated as a
linear function of its length. That is, we would estimate N2 as 2N1. Experimentally, we have found this
approximation to perform well despite density variations from nodes stopping at intersections. We defer methods for
learning each node’s exact N2 to future work.
2.2 Mathematical Analysis of Collision Rates
The general method of setting each node’s broadcast rate to 2N2t was shown in [1] to maximize network throughput
for arbitrarily high node densities; related cases of variable transmission distances are considered in [27]. At low
densities, however, other methods may approach 100% channel utilization with no collisions, e.g., a two-node
network where nodes alternate transmitting and receiving. We assume a network policy whereby all nodes share the
channel equally. We do not explore relaxing this assumption. However, nodes that are permitted to transmit more
often (e.g., if they generate more frequent updates) might simply be counted as multiple nodes in density estimates.
The proposition below shows the collision rate resulting from a 2N2t broadcast rate.
Collision Rate Proposition: For uniform density, two-hop node population N2, and same-size packets, the
broadcast rate 2N2t causes a collision rate that converges, as density increases, to 1/e, or about 36.8%.
Proof of the Collision Rate Proposition: Let the broadcast of any packet require time t. Consider a node n
broadcasting in the interval (0, t). A collision may occur if another node m within two hops of n begins its
broadcast in the interval (-t, t); especially at low densities, there might be no nodes in the shared transmission
area of m and n. If every node waits 2N2t between broadcasts, the probability of m causing a collision is at most
2t/2N2t = 1/N2. Let the set of possible collisions with n’s broadcast be a set of x events, each with probability
1
x−1
x
1/x. The probability of none of the x events occurring is (1 − )x = ( )x = ( )−x . As x increases, this
1
1
x
e
x
x
x−1
quantity approaches (1 + )−x = .
As CACL avoids collisions between one-hop neighbors, for any node n only those nodes in N2 but not in N1 can
cause collisions. Given same-size packets and uniform, high density, this collision avoidance limits the collision rate
to ((N2 – N1)/N2)(1/e). If N2 = 2N1, as for uniform node density on a road segment, the calculated collision rate is
18.4%. For uniform density and an open map (i.e., unconstrained by roads), N2 = 4N1 and the calculated collision
rate is 27.6%. Our results in section 3 will show that CACL does in fact approach these calculated collision rates.
Additionally, nodes broadcast almost immediately upon completion of a receipt (as was described in 2.1), which
tends to improve channel utilization.
This model assumes a collision at a node n prevents reception of all involved packets at n. Salvaging colliding
packets may be possible by techniques like multi-packet reception [37], [38]. This creates different issues – in the
extreme case, if all packets can be received despite collisions, why have a MAC layer? More likely, a fraction of
colliding packets could be received, reducing packet loss for both MAC models. CACL could then allow nodes a
faster broadcast rate to achieve the throughput-maximizing 1/e “true” collision rate (i.e., with packet loss).
3 Comparison of Collision Rates
Adopting CACL over the 802.11 MAC for broadcasting is advantageous when CACL offers a lower collision rate,
which is a function of both network traffic and topology. At a given node, maximum traffic simply means the
channel is constantly is use. However, a rather wide range of node densities can yield this maximum traffic level. To
prove this point, we provide the following two propositions. The first considers the minimum density of
broadcasting nodes required to ensure every point in a plane lies within range of at least one. The second considers
the maximum density of nodes that can broadcast simultaneously without causing any single-hop collisions.
Minimum Broadcast Density Proposition: For a uniform node density and transmission range r, the minimum
density of broadcasting nodes required to cover every point in a plane P is 2/3√3r2.
Proof of the Minimum Broadcast Density Proposition: Let the area within transmission range of a broadcasting
node be a circle of radius r. The coverage of P by such circles is a problem of tessellation by identical regular
polygons which are circumscribed by circles. It is well-known that such tessellation is only possible with triangles,
squares, or hexagons. Further, it is trivial to show that hexagons require the lowest density of circles to cover P and
we therefore omit this lemma.
If every node in a plane is either broadcasting or receiving, every point in P must be within at least one circle of
radius r whose centers are all at least r apart. To achieve this, tessellate the plane with regular hexagons with side
length r and area H. Let a circle of radius r and area Cr be circumscribed about each hexagon. These quantities are
shown in Fig. 2, below. Every point in P is within at least one circle of radius r. Since the distance from the center of
a hexagon to the midpoint of one of its sides is √3
𝑟, the centers of any two circles must be at least √3𝑟 apart. Since
2
the set of hexagons exactly covers the plane, the broadcast density (one node per circle of radius r) must be
2𝜋
multiplied by the ratio Cr/H = 3√3
≈ 1.21 and d is approximately 1.21
.
𝜋𝑟2
Fig. 2. Minimum Broadcast Density Tessellation
Fig. 3. Maximum Broadcast Density Tessellation
Maximum Broadcast Density Proposition: For a uniform node density and transmission range r, the
maximum density of broadcasting nodes without causing any single-hop collisions is 8/3√3r2.
Proof of the Maximum Broadcast Density Proposition: The minimum broadcast density from the previous
proposition can be increased without causing single-hop collisions so long as any two circles’ centers are
separated by at least distance r. To have the distance between circle centers approach r, tessellate the plane with
3√3 2
regular hexagons with side length √3
4 𝑟 and area 𝐻. Then, 𝐻 = 8 𝑟 . Let a circle with radius r and area Cr be
centered at the center of each hexagon, as shown in Fig. 3. Every point in P is within at least one circle of radius
r and all circles are at least r apart. Since the set of hexagons exactly covers the plane, the broadcast density of
8𝜋
one node per circle of radius r must be multiplied by the ratio Cr/𝐻= 3√3
≈ 4.84 and d is approximately 4.84
. This
𝜋𝑟2
density cannot be exceeded; moving the center of any circle C in any direction on P will bring it within less
than r of at least one other circle’s center.
To sidestep correlations between density and broadcast rate, our simulations vary each separately. Simulations were
run on the JiST/SWANS network simulator (available at (http://jist.ece.cornell.edu/). We assumed a 1km2 map,
using the 802.11b wireless standard (with bandwidth of 11Mbps and transmission range 50m). We used two
mobility models: an “open map” Random Direction model (which yields a constant node density), with node speeds
varying from 5 – 10m/s between direction changes, and a road-based vehicular traffic model, with vehicles moving
to random destinations over a rectangular grid of road segments spaced 100m apart, using the realistic STRAW
traffic module (available at http://www.aqualab.cs.northwestern.edu/projects/STRAW/index.php) and vehicle
speeds up to 23.4m/s.
Because of this rather dense road network and vehicle population, it was not possible to obtain an appropriate set of
real-world vehicular mobility data – all vehicles in the center of a large city. For example, the recent Microsoftsponsored project to track taxis in Beijing included 33,000 taxis but only a small fraction of total vehicular traffic
[44]. Similarly, many state and municipal transportation departments track vehicular traffic along major roads but
are not designed to collect the fine-grained information which our simulations require. Instead, we used the STRAW
module for the JiST/SWANS network simulator, which permits realistic micro-simulation of driver behaviors over
user-specified road networks.
We tested two primary independent variables to show the effects of increasing network traffic. The first was varying
the delay after broadcasts (only applicable to the 802.11 MAC, as CACL’s broadcast rate is density-driven). The
delay was randomly selected from a range with a minimum of 0.1ms and a maximum of 5 – 50ms, with a constant
node population of 600. The second was varying the node population, from 200 to 1800, with a maximum broadcast
delay (for the 802.11 MAC) of 5ms. To show the impact of varying packet sizes, we also tested two distributions:
uniform, all packets at 512 bytes, and mixed, with half the packets at 512 bytes and half at 50 bytes. A real-world
uniform distribution model would be push-based information dissemination, while the mixed distribution models
both request-response (with small query packets) and multiple applications with different packet sizes.
We present results for collision rates (i.e., packets actually received divided by the total possible received packets)
and total packets broadcast. Note that, since CACL adapts the broadcast rate to channel capacity, these results are
constant for changing broadcast delays within each packet size distribution. We also present results for adjusted
reception difference (ARD), which considers the combined effects of collision rates and packets broadcast to
estimate the relative likelihoods of receipt for CACL and the 802.11 MAC. We define ARD as follows. For a given
set of parameters, let the collision rate for CACL be c and packets broadcast be b and the same variables for the
′
′
802.11 MAC be c’ and b’. Then, the ARD is (1 − 𝑐 ′ )(𝑏 /𝑏) − (1 − 𝑐)𝑏/𝑏 = (1 − 𝑐 ′ )(𝑏 /𝑏) − (1 − 𝑐). The left term is
the probability of successful receipt for (𝑏 ′ /𝑏) broadcasts. By subtracting (1 − 𝑐), we obtain the probability of
receipt for a single broadcast under CACL, i.e., the difference ARD. In short, a positive ARD means the 802.11
MAC should offer higher throughput, while a negative ARD means CACL does.
An examination of Figs. 4 through 7 shows two main patterns. First, for the 802.11 MAC there is a strong
correlation between the collision rate and the number of packets broadcast. This shows that, at high broadcast rates,
CACL is less collision-prone than the 802.11 MAC. Second, the mixed packet distribution yields lower collision
rates for the 802.11 MAC (due to smaller packets being broadcast with the same average delay afterward) but not
for CACL. In Figs. 5 and 7, the packet counts for both MAC models are nearly identical and so the mixed
distribution results are obscured by the uniform distribution. Collision rates for both CACL sets are below the
calculated collision rates of 27.6% for the open map and 18.4% for road-based mobility (the discrepancy is due to
lower density near the map boundary), showing CACL’s robustness against packet size variations.
Fig. 4. Collision rates by broadcast rate, open map
Fig. 5. Packets broadcast by broadcast rate, open map
Fig. 6. Collision rates by broadcast rate, roads
Fig. 7. Packets broadcast by broadcast rate, roads
Results for varying node populations are given in Figs. 8 through 11. Here, the collision rate sharply increases for
the 802.11 MAC as density increases. For CACL, the collision rate is stable and very close to the calculated rate on
the open map (Fig. 8). For road-based mobility, CACL’s collision rate increases slowly and exceeds the calculated
rate (Fig. 10); the reason is that, as density increases, intersections become congested which reduces the reliability of
density estimates. Packets broadcast increase almost linearly with density for the 802.11 MAC (Figs. 9 and 11),
while they are essentially stable for CACL on the open map (Fig. 9) and increase only slowly for road-based
mobility (Fig. 11). These results show how CACL promotes stable network performance despite large variations in
node density.
Fig. 8. Collision rates by node population, open map
Fig. 9. Packets broadcast by node population, open map
Fig. 10. Collision rates by node population, roads
Fig. 11. Packets broadcast by node population, roads
Fig. 12. Adjusted reception difference
by broadcast delay
Fig. 13. Adjusted reception difference
by node population
Let us now consider throughput, as shown by ARD. CACL does outperform the 802.11 MAC over most of the
tested delays (Fig. 12) and node populations (Fig. 13). However, at the lowest tested average delays, the throughput
of the 802.11 MAC in fact exceeds that of CACL. This is due to the node density; while CACL is designed to
maintain constant throughput for increasing densities, the same is not true of the 802.11 MAC. This is shown by
increasing density while keeping the delays constant. Collisions for the 802.11 MAC increase substantially with a
net negative impact on throughput. It should, however, be noted that for road-based mobility and uniform packet
sizes, the 802.11 MAC does provide higher throughput than CACL for low to moderate node densities. On the other
hand, CACL is very consistently superior to the 802.11 MAC for cases of mixed packet sizes, and greatly so as
density increases. This shows that CACL is well-suited for applications that are broadcast-reliant and permit
variable packet sizes. Perhaps most importantly, CACL can simultaneously support a variety of applications which
each uses a potentially different range of packet sizes.
4 Multi-Hop Routing: Geographically-Constrained Flooding
In this section, we compare CACL against CACL-G in a multi-hop vehicular ad hoc network scenario. Single-hop
reception is an important but limited metric for MAC performance. In the previous section, we considered network
traffic, node density, and different mobility models. All else equal, we would expect a low broadcast collision rate to
be correlated with improved performance in multi-hop routing protocols. Most importantly, fewer transmissions
would be needed (on average) to forward each packet to the next node. This would enable higher network
throughput and faster end-to-end delivery. It would also make the network more robust against increasing traffic
levels, as packets would be less prone to being dropped while being routed to their destinations.
However, multi-hop routing adds additional variables that complicate assessing the performance of medium access
controls. First, medium access controls must operate in conjunction with a routing protocol. Different applications,
routing protocols, and scenarios may be more or less reliant on broadcasting. Notably, higher mobility tends to
require more frequent broadcasting, whether to learn routing paths or because packets are flooded. Multi-hop routing
is further affected by the potential availability of numerous possible routing paths, each of which may have
substantially different characteristics. For example, if two disjoint paths exist, one favorable to CACL and the other
favorable to the 802.11 MAC, the two medium access controls may both perform equally well but only while both
paths are available.
In vehicular applications, multi-hop routing is largely confined to roads. Compared to random walk mobility
models, where nodes may move in any direction, this confinement to roads has two primary consequences. First, for
a given node population, the effective node density is higher. That is, nodes will on average have more one-hop
neighbors. Second, there are fewer possibilities for bypassing any large gaps (greater than communication range)
between nodes. If such a gap arises on a road segment, it would prevent all routing until nodes move within mutual
communication range. The net effect of these two consequences is difficult to predict. Road-based mobility models
more often find complete end-to-end routing paths at lower node populations but path breakage is highly correlated.
For these simulations we again used the JiST/SWANS network simulator with the STRAW transportation module.
We largely reused the road network model from section 3 but assigned vehicles to specific routes. For the initial
placement, vehicles were assigned to roads using a uniform probability distribution. For route choice, each vehicle
moves either in a clockwise or counter-clockwise square loop (both equally likely) in a 500m square. A vehicle
starting at A (Fig. 14) might follow the path A-B-C-D-A or A-D-C-B-A. When equal numbers of vehicles are
assigned to each square route, we achieve uniform vehicle density over the road network as well as in regard to
traffic directions on each road segment. We present results for three uniform density scenarios and one non-uniform
density scenario. Of course, real-world vehicular traffic does not exhibit uniform density but such results can be
extrapolated to estimate network performance under non-uniform densities.
Fig. 14. Travel route example.
As in the previous section, we have again used the 802.11b wireless standard instead of the 802.11p standard, which
was designed expressly for vehicular applications. First, its long transmission range (typically modeled as 300m to
1000m, compared to 50m for 802.11b) is not desirable for the high vehicle densities which often arise in urban
areas. Compounding this, the limited bandwidth of 802.11p makes it not well-suited for multi-hop routing. The two
basic solutions to these characteristics of 802.11p are limiting each node’s transmission rate, as in [40], [12], and
limiting transmission range, as in [43]. We are, in fact, implementing these very two solutions by our selection of
CACL and the 802.11b wireless standard.
Our paper [25] evaluated CACL’s performance in regard to a vehicular ad hoc network application based on multihop routing using the SBSD controlled flooding routing protocol. We observed that CACL outperformed the 802.11
MAC at all but the lowest tested network traffic levels. Here, we extend this combination of CACL and SBSD by
incorporating a spatial constraint extension, CACL-G. As we have already published results comparing CACL and
the 802.11 MAC and because the 802.11 MAC’s performance is further strongly dependent on the redundant
transmission rate, we only compare CACL against CACL-G. Within this framework, we test the effects of varying
node density while maintaining a fixed query posting rate, with details given below:
Node population: We include results for 600, 900, and 1200 vehicular nodes, with all vehicles assumed to
participate in the network. This variable node population has a fixed set of 150 query sources. The lower bound of
600 nodes was chosen for being the minimum density where most queries are able to receive responses at the
routing distances consequent to our query-response model, which follows.
Queries and responses: Queries seek location-based information along the source’s travel route. Because our
focus is on routing performance, our query model is abstract; readers interested in more detailed consideration of
publish-subscribe or context-management are directed to [20] and [13], respectively. Referring to Fig. 14, each
vehicle follows a loop from its starting point A (A-B-C-D-A or A-D-C-B-A). A source’s query relates to its soonest
upcoming A or C waypoint and may be answered by any vehicle within 50m of the corresponding intersection (in
rare cases, a source may immediately answer its own query). This simulates gathering location-specific information,
such as travel conditions on the adjacent roads. As routing must conform to the road network, the effective routing
distance is up to 1000m. Each query packet has size 50 bytes and each response packet has 512 bytes. Each query
source posts queries at an average interval of 5s and with lifetime 10s. Importantly, vehicular travel during a query’s
lifetime is much shorter than the expected routing distance; this minimizes the impact of any potential differences in
travel route selection from our mobility model.
Geographical constraint: The basic constraint is a square S with the query source’s current location at one
corner and the query destination at the opposite corner. Under CACL-G, a vehicle v will only receive or transmit a
packet p with corresponding constraint area S(p) if v is itself within A(p). We observe the effects of expanding this
boundary on all sides by 50m to 450m; the higher range allows most packets to be flooded over nearly the entire
map, which produces performance similar to that of the basic CACL model.
Simulation runs: Each run proceeds for 40s of simulated time. Data are presented for the middle 20s of each
run, which allows 10s at the start for the network to reach equilibrium traffic and 10s at the end for any included
queries to either receive responses or expire. Each data point represents the average of ten simulation runs. Note
however, although the CACL results are presented as a line with markers, only one set of CACL runs was performed
for each node population level; the boundary extension does not affect the basic CACL model.
Results are presented for four performance variables:
Delivery: The percentage of queries which receive a matching response within the query’s lifetime.
Response time: The average time for a query to receive a matching response.
Query packets transmitted: The total number of query packets transmitted (by all nodes)
Response packets transmitted: The total number of response packets transmitted (by all nodes)
The results for 600 nodes are presented in Figs. 15 – 18. With a relatively tight geographical constraint, delivery
(Fig. 15) is markedly improved, with an increase of as much as 10%. At the tightest tested constraint, response time
(Fig. 16) is also improved due to the reduced network traffic within the flooding area for each query and response.
For the more relaxed constraints, response time is similar to that of CACL. For packet counts (Figs. 17 and 18),
tighter constraints lead to a greater proportion of network traffic being devoted to responses. Since queries are more
closely routed to their destinations, this is expected. However, at the tightest constraint, both query and response
counts decrease for CACL-G. This happens because nodes near the map corners and edges occasionally have no
packets that are eligible for broadcast due to the restrictions of CACL-G.
Fig. 15. Delivery, 600 nodes
Fig. 16. Response time, 600 nodes
Fig. 17. Query packets transmitted, 600 nodes
Fig. 18. Response packets transmitted, 600 nodes
The results for 900 nodes are presented in Figs. 19 – 22. Delivery rates (Fig. 19) show greater improvement of
CACL-G over CACL, with a gain of almost 20% for the 150m boundary extension case. It is noteworthy that this is
almost entirely due to the worsening performance of the basic CACL model. Without an additional constraint, each
packet is transmitted by more nodes, which wastes bandwidth as density increases. Response time (Fig. 20) is nearly
identical for both CACL and CACL-G. A small improvement is, however, seen at the 350m boundary extension.
This is from two particularly favorable simulation runs for CACL-G but also because the worsening delivery leads
to fewer distant responses being found and delivered.
For packet counts (Figs. 21 and 22), the same general trends are observed as in Figs. 17 and 18. However, because
of the increased node density, the number of query transmissions is lower overall and the number of responses is
higher. This happens because, with increasing density, a query is more likely to have a continuous path to its
destination and travel over fewer hops to reach it. Also, we again see that tighter constraints lead to fewer query
transmissions but more responses. This phenomenon also affects the network traffic handled by isolated nodes at the
map corners. At the tightest constraint, query and response transmissions (Figs. 21 and 22) are both lower for
CACL-G than for CACL but (compared to Figs. 17 and 18) are dominated by the reduction in queries.
Fig. 19. Delivery, 900 nodes
Fig. 20. Response time, 900 nodes
Fig. 21. Query packets transmitted, 900 nodes
Fig. 22. Response packets transmitted, 900 nodes
The results for 1200 nodes are presented in Figs. 23 – 26. Here, we see different patterns begin to emerge from the
higher density. First, CACL-G outperforms CACL in terms of delivery (Fig. 23) at all constraint levels. This is
largely due to the declining performance of CACL without constraints on redundancy. Response time, however,
tends to be somewhat slower for CACL-G than CACL; this is a consequence of the lower delivery rate. As well, in
this particular scenario, the declining delivery rate for CACL-G as the boundary is extended is reflected in a stable
response time.
Query counts (Fig. 25) are lower for all tested boundaries and dramatically so at the tightest constraints. This occurs
because of the prevalence of continuous end-to-end routing paths. Most queries are able to find a matching response
and do so relatively quickly. Thus, network traffic is dominated by responses. However, because responses are so
much larger than queries, the increase in response packet counts is only noticeable at the 150m boundary extension
and, even then, is not especially large.
Fig. 23. Delivery, 1200 nodes
Fig. 24. Response time, 1200 nodes
Fig. 25. Query packets transmitted, 1200 nodes
Fig. 26. Response packets transmitted, 1200 nodes
In our last set of results, we consider a non-uniform distribution of vehicular traffic. The roads shown as heavy lines
(Fig. 27) have twice as many vehicles as the others. This mobility again uses the square routes shown in Fig. 14; in
effect, the heavy traffic has four square routes, each starting at one of the map corners. On the heavy-traffic roads,
vehicle density is comparable to the 1200 node scenario; on the others, it is comparable to the 600 node scenario.
Intuitively, the results for this non-uniform mobility model should exhibit trends corresponding to an intermediate
density between the light and heavy traffic. In fact, we do observe very similar trends (Figs. 28 – 31) to those from
the uniform density scenarios.
Fig. 27. Non-uniform road density
Fig. 28. Delivery, 900 nodes, non-uniform density
Fig. 29. Response time, 900 nodes, non-uniform density
Fig. 30. Query packets transmitted, 900 nodes,
non-uniform density
Fig. 31. Response packets transmitted, 900 nodes,
non-uniform density
The overall picture given by these results is that, as density increases, any geographical constraint on flooding
appears beneficial. However, the best results are achieved by including a slightly wider set of roads than only the
most direct path. A boundary of 150m, which offered the best delivery in all four scenarios, includes one adjacent
road surrounding the square defined by the source and destination. For real-world implementations, density could be
assessed or estimated along potential routing paths or else vehicles could adapt their transmission strengths to
facilitate routing along relatively low-density road segments.
5 Related Research
A good recent survey of cross-layer approaches is provided by [14], grouping cross-layer methods by which layers
are involved. Ultimately, cross-layer design might combine the entire OSI stack into a single optimized module [2].
Although cross-layer methods can improve system performance, they also increase system complexity and designers
should be cautious [15]; in that regard, we note CACL is structurally and computationally simple. For contextual
factors, methods to regulate packet forwarding have been proposed using channel capacity [21], contention [9],
signal fading [36], [41], battery power [42], and topology [8]. These methods, however, are either computationally
onerous or require gathering information not already available to the 802.11 MAC.
Thorough surveys of medium access controls are provided by [19] and [26], the latter focusing on their usage in
VANETs. While many researchers have proposed MACs specifically for VANETs, they are quite often specifically
designed for DSRC and, unlike CACL, there is no evidence they will perform well for 802.11a/b/g/n wireless
standards. For example, a simple method for increasing contention windows based on the rate of successful packet
receipts is presented in [5]. However, their approach is intended for relatively long transmission ranges, so that the
pool of data points remains large and covers a long-lasting set of vehicles.
A method for location-based broadcast scheduling (toward providing VANETs fair and reliable access to roadside
Internet stations) is given in [18]. In essence, each road is divided into segments and vehicles in adjacent segments
do not transmit simultaneously. The practical implementation issues are variability of transmission range (as
segments are pre-designated), density variations across segments, and the potential unreliability of position
estimates. Further, this approach is intended to regulate data traffic flows around static infrastructure points. No
solution is proposed for cases of diverse, highly mobile message destinations.
A method for scheduling transmissions within slots is proposed in [4]. This method does not require a priori
knowledge of node density or population to converge to good performance, as it learns local conditions and adapts
to them. However, in dynamic environments where those parameters are constantly shifting, there is no assurance it
will perform well. CACL, on the other hand, is robust against density changes.
A context-aware MAC for vehicular networks is proposed in [35]. Their model combines recent link quality history
with a predictive model including vehicular positions, speeds, and accelerations. While their model performs well
for regulating communications between vehicle pairs, it is not well-suited for broadcasting. In dense environments,
frequently gathering detailed information from every nearby vehicle imposes a substantial overhead burden. It is
also unclear how to resolve cases of multiple recipients with greatly differing contextual states.
A method for broadcasting in VANETs is described in [39]. Nodes evaluate whether their neighborhood is dense, in
which case rebroadcasts are generally avoided, or sparse, in which case “store-carry-forward” methods are adopted.
However, the results are not readily compared to delay-tolerant approaches, which inevitably include multiple
transmissions. The issue of delay-tolerance is orthogonal to the broadcast rate in CACL, which is equally suited to
single or multiple transmissions of a given packet.
One of the early approaches to geographically-constrained flooding was the seminal paper [17]. Its proposed
Location-Aided Routing (LAR) essentially defined flooding areas as polygons including the source and destination.
More recent approaches, such as [33], use a method termed mobility-centric geocasting. The approach is to estimate
node density from mobility information, as vehicular traffic tends to slow as density increases.
6 Conclusion
We have presented CACL, a context-aware broadcasting scheduling model. CACL addresses an important limitation
of the 802.11 MAC for broadcast-reliant applications: providing consistent performance under heavy network loads.
We have tested CACL for single-hop collisions and shown that it substantially reduces the broadcast collision rate
compared to the 802.11 MAC. Although the 802.11 MAC offers superior performance under low-traffic conditions,
adopting CACL becomes increasingly beneficial as node density and network traffic increase.
We have also presented a geographically-constrained extension to CACL, CACL-G. This model improves on the
performance of CACL in multi-hop routing applications. Our simulation results showed that, as node density
increases, a more restrictive flooding policy can improve end-to-end routing performance. This happens because
higher density increases the likelihood of continuous end-to-end routing paths; restricting flooding then has little
effect on reachability. Moreover, high density makes redundant transmissions more prevalent, making restrictions
necessary to maintain network throughput.
Our results have shown boundaries where CACL outperforms the 802.11 MAC, where it does not, and where the
results are unclear. These boundaries suggest a decision function where nodes might broadcast using CACL or the
802.11 MAC. Likewise, at low node densities, the best policy is no geographic restriction at all; the restrictions of
CACL-G are superfluous. Ultimately, no one MAC is optimal in all possible network conditions and, therefore,
adaptive hybrid protocols are promising. We are pursuing several methods for improving CACL, both in gathering
the density information already used and in learning about different contextual variables.
Our future work on the basic CACL model includes tracking two-hop node neighborhoods, coordinating broadcast
schedules to reduce collisions (when node groups have relatively stable topology), and scheduling sets of broadcasts
according to packet size (to reduce the negative impact of any collisions that do occur). For CACL-G, we are
exploring sampling methods for learning density along road segments in order to optimize the geographical
constraints. An alternative strategy is adjusting transmission range to compensate for low density; however, the
consequent additional hidden terminal collisions and reductions in the wireless communication data rate present
difficult challenges.
Acknowledgments
This work is supported by a University of Illinois College of Business Administration grant and National Science
Foundation grant CNS-0910988.
References
1. Abramson, N., “The Throughput of Packet Broadcasting Channels,” IEEE Transactions on Communications, vol.
25, pp. 117–128, 1977.
2. Akyildiz, I.,Vuran, M., and Akan, Ö., “A Cross-Layer Protocol for Wireless Sensor Networks,” in Proc. CISS,
2006.
3. Akyildiz, I. and Wang, X., “A Survey on Wireless Mesh Networks,” IEEE Communications Magazine, vol. 43,
no. 9, s23-30, September, 2005.
4. Baccelli, F., Blaszcyszyn, B., and Muhlethaler, P., “An ALOHA Protocol for Multihop Mobile Wireless
Networks,” IEEE Transactions on Information Theory, vol. 52, pp. 421–436, 2006.
5. Balon, N. and Guo, J., “Increasing Broadcast Reliability in Vehicular Ad Hoc Networks,” in Proc. VANET, 2006.
6. Bianchi, G., “Performance Analysis of the IEEE 802.11 Distributed Coordination Function,” IEEE Journal on
Selected Areas in Communications, vol. 18, pp. 535-547, 2000.
7. Brenner, P., “A Technical Tutorial on the IEEE 802.11 Protocol,” BreezeCom Wireless Communications, 1997.
8. Canli, T., Nait-Abdesselam, F., and Khokhar, A., “A Cross-Layer Optimization Approach for Efficient Data
Gathering in Wireless Sensor Networks,” in Proc. INCC, 2008.
9. Chen, L., Low, S., Chiang, M., and Doyle, J., ‘Cross-Layer Congestion Control, Routing and Scheduling Design
in Ad Hoc Wireless Networks,” in Proc. IEEE INFOCOM, 2006.
10. Choi, J., So, J., and Ko, Y., ‘Numerical Analysis of IEEE 802.11 Broadcast Scheme in Multihop Wireless Ad
Hoc Networks,” Information Networks, Springer Verlag, 2005.
11. Ekambaram, V. and Ramchandran, K., “R-GPS (Robust GPS): Enhancing GPS Accuracy and Security using
DSRC,” ITS World Congress: Special Session for Connected Vehicular Technology Challenge Winners, 2011.
12. Ghafoor, K., Bakar, K., van Eenennaam, M., Khokhar, R., and Gonzalez, A., “A Fuzzy Logic Approach to
Beaconing for Vehicular Ad Hoc Networks,” Telecommunication Systems, vol. 47, pp. 1–11, 2011.
13. Jaroucheh, Z. and Liu, X., “An approach to domain-based scalable context management architecture in pervasive
environments,” Journal of Personal and Ubiquitous Computing, vol. 16, pp. 741–755, 2012.
14. Jarupan, B. and Ekici, E., “A survey of cross-layer design for VANETs,” Ad Hoc Networks, vol. 9, pp. 966–983,
2011.
15. Kawadia, V. and Kumar, P., “A Cautionary Perspective on Cross Layer Design,” IEEE Wireless
Communications, vol. 12, pp. 3–11, 2005.
16. Khurana, S., Kahol, A., and Jayasumana, A., “Effect of Hidden Terminals on the Performance of IEEE 802.11
MAC Protocol,” IEEE LCN’98, 1998.
17. Ko, Y. and Vaidya, N., “Location-Aided Routing (LAR) in mobile ad hoc networks,” Wireless Networks, vol. 6,
pp. 307-321, 2000.
18. Korkmaz, G., Ekici, E., and Ozguner, F., “A Cross-Layer Multihop Data Delivery Protocol with Fairness
Guarantees for Vehicular Networks,” IEEE Transactions on Vehicular Technology, vol. 55, pp. 865–875, 2006.
19. Kumar, S., Raghavan, V., and Deng, J., “Medium Access Control protocols for ad hoc wireless networks: A
survey,” Ad Hoc Networks, vol. 4, pp. 326–358, 2006.
20. Lin, C., Jin, B., Long, Z., and Che, H., “On context-aware distributed event dissemination,” Journal of Personal
and Ubiquitous Computing, vol. 15, pp. 305–314, 2011.
21. Lin, X. and Shroff, N., “The Impact of Imperfect Scheduling on Cross-Layer Congestion Control in Wireless
Networks,” IEEE Transactions on Networking, vol. 14, pp. 302–315, 2006.
22. Lundquist, D. and Ouksel, A., “An Efficient Demand-Driven and Density-Controlled Publish/Subscribe Protocol
for Mobile Environments,” Inaugural Int’l Conference on Distributed Event-Based Systems (DEBS 2007), Toronto,
Canada, June, 2007.
23. Lundquist, D. and Ouksel, A., “Dynamic Subscription Permission: Extending the Depth of Demand-Controlled
Flooding,”Proc. 2008 IEEE Asia-Pacific Services Computing Conference (APSCC 2008), Yilan, Taiwan, December,
2008.
24. Lundquist, D. and Ouksel, A., “Distributed Delay: Improving Network Throughput by Reducing Temporal
Saturation,” NTMS, 2009.
25. Lundquist, D. and Ouksel, A., “A Context-Aware Cross-Layer Broadcast Model for Ad Hoc Networks: Analysis
of Multi-Hop Routing,” MobiWIS, 2012.
26. Menouar, H., Filali, F., and Lenardi, M., “A Survey and Quantitative Analysis of MAC Protocols for Vehicular
Ad Hoc Networks,” IEEE Wireless Communications, vol. 13, pp. 30–35, 2006.
27. Metzner, J., “On Improving Utilization in ALOHA Networks,” IEEE Transactions on Communications, vol. 24,
pp. 447–448, 1976.
28. Oliveira, R., Bernardo, L., and Pinto, P., “The influence of broadcast traffic on IEEE 802.11 DCF networks,”
Computer Communications, vol. 32, pp. 439–452, 2009.
29. Ouksel, A. and Lundquist, D., “Demand-Driven Publish/Subscribe in Mobile Environments,” Wireless Networks
vol. 16, no. 8, pp. 2237–2261, 2010
30. Ouksel, A. and Lundquist, D., “A Context-Aware Cross-Layer Broadcast Model for Ad Hoc Networks,” ANT
2012, 2012.
31. Ouksel, A. and Lundquist, D., “Pollution Credit Trading in Vehicular Ad Hoc Networks,” ITS World Congress:
Special Session for Connected Vehicular Technology Challenge Winners, 2011.
32. Peng, C., Tan, Y., Xiong, N., Yang, L., Park, J., and Kim, S., “Adaptive video-on-demand broadcasting in
ubiquitous environment,” Journal of Personal and Ubiquitous Computing, vol. 13, pp. 479–488, 2009.
33. Piorkowski, M., “Mobility-Centric Geocasting for Mobile Partitioned Networks,” in Proc. IEEE International
Conference on Network Protocols, pp. 228–237, 2008.
34. Resta, G, Santi, P., and Simon, J., “Analysis of multi-hop emergency message propagation,” in Proc. ACM
MobiHoc, 2007.
35. Shankar, P., Nadeem, T., Rosca, J., and Iftode, L., “CARS: Context-Aware Rate Selection for Vehicular
Networks,” in Proc. ICNP, 2008.
36. Song, G. and Li, Y., “Cross-Layer Optimization for OFDM Wireless Networks–Part I: Theoretical Framework,”
IEEE Transactions on Wireless Communications, vol. 4, pp. 614–624, 2005.
37. Tong, L., Zhao, Q., and Mergen, G., “Multipacket Reception in Random Access Wireless Networks: From
Signal Processing to Optimal Medium Access Control,” IEEE Communications Magazine, vol. 39, pp. 108 – 112,
2001.
38. Tong, L., Naware, V., and Venkitasubramaniam, P., “Signal Processing in Random Access,” IEEE Signal
Processing Magazine, vol. 21, pp. 29-39, 2004.
39. Tonguz, O., Wisitpongphan, N., Bai, F., Mudalige, P., and Sadekar, V., “Broadcasting in VANET,” in Proc.
INFOCOM, 2008.
40. van Eenennaam, M., Wolterink, W., Karagiannis, G., and Heijenk, G., “Towards Scalable Beaconing in
VANETs,” Fourth ERCIM Workshop on e-Mobility, 2010
41. Viswanath, P., Tse, D., and Laroia, R., “Opportunistic Beamforming Using Dumb Antennas,” IEEE
Transactions on Information Theory, vol. 48, pp. 1277–1294, 2002.
42. Wu., Y., Chou, P., Zhang, Q., Jain, K., Zhu, W., and Kung, S., “Network Planning in Wireless Ad Hoc
Networks: A Cross-layer Approach,” IEEE Journal on Selected Areas in Communications, vol. 23, pp. 136–150,
2005.
43. Yang, L., Guo, J., and Wu, Y., “Channel Adaptive One Hop Broadcasting for VANETs,” 11th International
IEEE Conference on Intelligent Transportation Systems, pp. 369 – 374, 2008.
44. Yuan, J., Zheng, Y., Zhang, C., Xie, W., Xie, X., and Huang, Y., “T-Drive: Driving Directions Based on Taxi
Trajectories,” ACM SIGSPATIAL GIS 2010, pp. 99 – 108, 2010.