TRAC: An Architecture for Real-Time Dissemination ofVehicular Traffic Information

Shravan Rayanchu, Sulabh Agrawal, Arunesh Mishra, Suman Banerjee,∗

S. Ganguly†

{shravan, sulabh, arunesh, suman}@cs.wisc.edu, samrat@nec-labs.com

IntroductionTraffic congestion is a major cause of concern in many cities.A real-time information dissemination system announcingcurrent traffic conditions to travelers is increasingly becom-ing a necessity in this context. The recent DSRC researchinitiative[1] envisions a deployment of DSRC access points(APs) along the road side to enable Vehicle to RoadsideCommunication (VRC) for this purpose. However, even ifsuch an infrastructure is deployed on the roadside, it is nottrivial to design a system that distributes on-demand trafficinformation. Existing solutions employ a centralized designwhere the traffic information is stored at a central databaseand the vehicles then query the database for information oftheir interest (e.g. what is the average speeds on a set ofroads?). Such solutions are simply not scalable due to thehuge volume of queries the central database would have tohandle. Moreover, such a communication architecture doesnot take into account the fact that traffic related queries arelocation sensitive and that the sources for a particular queryalso tend to belocalized.

Publish/subscribe systems are considered to be very ef-fective for supporting data-centric, information dissemina-tion applications including traffic notifications[2]. Under apub/sub model, the vehicles could act as publishers as wellas subscribers. In the role of publishers, the vehicles publishtheir current road and speed information to an AP (wheneverunder the range) and as subscribers they query for informa-tion about certain road segments of their interest. The APscan then form a broker network to carry out pub/sub opera-tions. However, traditional pub/sub paradigm is inadequatefor the purposes of traffic information dissemination for anumber of reasons: (1) Traditional pub/sub systems assumethat the number of subscribers typically outnumber the pub-lishers. It is this assumption that justifies flooding of adver-tisements by each publisher. But in our system the numberof publishers and subscribers arelarge andequal in number.(2) Since both publishers and subscribers aremobile, the sys-tem would be highly inefficient under traditional pub/sub se-mantics. (3) Given the intermittent connectivity of the pub-lishers, the overhead of constructing a multicast tree rootedat each publisher would far outweigh any savings achievedin the delivery of publications. (4) Data published by vehi-

∗Department of Computer Science, University of Wisconsin Madi-son, USA†NEC Laboratories, NJ, USA

Broker for Region X

LWA

LWA

Broker

Publish/Subscribe Broker Network

Subscriber (static)

LWA

Publisher (Virtual Publisher for R1)

Subscriber

(1) Published tuples from vehicle

(2) Advertisements from Broker

(3) Subscriptions from vehicle

(4) Publications from Broker (virtual publisher)

DSRC access point

Geographical Region X Geographical

Region Y

(road

R1,

spee

d S1)

(road

R1,

spee

d S1)

(road

R1,

?)

Data centric routing

(1)

(2)

(3)

(4)

Figure 1: TRAC : Architecture

cles under the range of an AP would be similar in content.Exploiting this locality of publications by reusing the multi-cast trees would reduce the cost of mobility, but traditionalpub/sub system would be oblivious of such a property.

In this paper, we introduce TRAC, a scalable architecturefor vehicular traffic information dissemination based on apub/sub model. TRAC efficiently handles the above statedinadequacies using the novel concept ofvirtual publisherand optimizations likesubscription predictions. Our objec-tive while designing TRAC has been to distribute on-demandtraffic information to highly mobile users at a minimal la-tency (multicast tree building time) while incurring a rela-tively low message overhead.TRAC: Architecture & AlgorithmsBasic Idea: The primary drawback of traditional pub/subsystems is that they store state per publisher. Even whenmultiple publishers publish similar information, the systemconstructs a tree rooted at each publisher. We propose usingastateless method of publishing where the system employs atree per each data item rather than a tree per publisher. Whena publisher connects to a broker, it simply publishes the dataand disconnects. The broker then acts as a virtual publisheron behalf of the actual publisher by sending out covering ad-vertisements, building the multicast trees and publishingthedata. Since there can be multiple sources for the same data, ifa tree for the data already exists in the system, a new branchcreated at the broker grafts onto the tree opportunistically,else a new tree for the data is constructed before sending outthe publications. The concept of virtual publisher resultsinreduced (un)advertisement costs, as the trees are not (re)builtfor every movein/moveout operations of a publisher. More

50 150

Ave

rage

Tre

e B

uild

ing

Tim

e (m

s)

No of Subscribers

5

10

15

20

25

30 STANDARD-VP

PRESUB-NBR

PRESUB-PRED

250 350

Figure 2: Avg. Tree Building Time(250 Subs)

50 70 90

Ave

rage

Tre

e B

uild

ing

Tim

e (m

s)

Speed (mph)

5

10

15

20

25

30

STANDARD-VP

PRESUB-NBR

PRESUB-PRED

Figure 3: Avg. Tree Building Time(Speed=70mph)

STANDARD-VP

PRESUB-NBR

PRESUB-PRED

Mobility Scheme Average cost per

subscription

1

0.71

0.43

Figure 4: Average Cost perSubscription (normalized wrtSTANDARD-VP)

importantly, this indirection provided by the system reducesthe problem to the case of static publishers and mobile sub-scribers.

Architecture: In TRAC, the service area is divided intoa number of local regions, each of which is assigned a bro-ker. The vehicles are equipped with GPS, on-board diagnos-tics system (OBD) interface that can be used to collect data(speed of the vehicle etc) and a wireless network interface(802.11 or DSRC) to communicate with the APs. The ve-hicles periodically record the tupes(road id, speed)and publish whenever under coverage. A Light Weight Agent(LWA) resides on the DSRC/802.11 APs to which the pub-lishers/subscribers connect and participate in the publish /sub-scribe operation. The LWA is responsible for storing theconnection state, caching the publications and maintainingthe subscription tables. It acts as a relay agent and passeson the data to the corresponding broker for its region. As inany pub/sub system the broker is responsible for routing theadvertisements and publications, subscription matching andbuilding the multicast trees. Additionally, the broker acts asa virtual publisher by periodically publishing the aggregateinformation received from the LWAs in its region. We callthis proposed routing scheme theSTANDARD-VP scheme.Figure 1 shows the overall architecture.

Optimizations: In order to reduce the (un)subscriptioncosts arising due to subscriber mobility we used the follow-ing schemes: (1)PRESUB-NBR: Since there is a high possi-bility that a subscriber after moving out of an AP connectsto one of the APs in the neighboring area, we exploit thisproperty by sending the pre-subscriptions to all the neigh-boring APs (LWAs) within some radius. The latency as per-ceived by the user would then be minimal in this case. If asubscriber does not connect to a pre-subscribed AP, unsub-scriptions are issued after a timeout. (2)PRESUB-PRED: Inthis scheme, an LWA upon receiving an unsubscription pre-dicts the next neighboring AP based on the road segments inthe subscribed path and sends out the pre-subscription. Thiswould require more processing on the part of an LWA, but itresults in a lesser message overhead.SimulationsWe developed the TRAC Simulator by enhancing GrooveSim[3]and integrating it with a discrete event based simulator whichmodeled a pub/sub system. GrooveSim generates a graphbased abstraction of streets for any place in the USA by

importing the TIGER/Line files. Communication betweenthe vehicles, APs and the brokers is modeled using sim-ple message queues taking appropriate latencies (2ms foreach link) into account. We simulated our experiments usingthe map of Manhattan county (service area of23.7 squaremiles divided into16 regions with a total of9312 road seg-ments). A total of92 APs (LWAs) were placed at road in-tersections, uniformly distributed with a mean distance ofaround600 meters. We initialized650 vehicles at randomaddresses in the area which subscribed to a randomly cho-sen path whenever under coverage. Figure 2 shows the av-erage multicast tree building time for increasing number ofsubscribers. Figure 3 shows the same for increasing mo-bility. The STANDARD-VP scheme performs very well witha tree building time of less than25ms. Since the averageconnection time even at90mph was about5 seconds, allthe publications were delivered to the interested subscribers.The PRESUB-PRED and PRESUB-NBR schemes further re-duce the tree building time. Moreover, the tree building timeis impervious to increasing number of subscribers/mobilitywhich exhibits good scalability properties of the architec-ture. The average message overhead forSTANDARD-VP schemewas 7.24 messages/subscription. This low message over-head can be attributed to the concept of virtual publisherwhich results in reduced (un)advertisement costs. Figure 4shows the normalized cost. As before, the proposed opti-mizations perform better than theSTANDARD-VP scheme.Conclusion and Future WorkWe presented TRAC: a scalable, distributed traffic informa-tion dissemination system. TRAC uses the concept of vir-tual publishers and subscription predictions to completelydecouple publisher mobility from the system and enable traf-fic information aggregation from multiple publishers. Weare currently investigating the practical issues concerningthe throughput and range of coverage with drive-thru ex-periments over 802.11b. We further plan to investigate theimpact of different broker network topologies on the perfor-mance of TRAC.References[1] DSRC WG, http://grouper.ieee.org/groups/scc32/dsrc/index.html

[2] V.Muthusamy, A. Jacobsen, ”Effects of routing computations in content-

based routing networks with mobile data sources”, MOBICOM’05

[3] R. Mangharam, R. Rajkumar, ”GrooveSim: a topography-accurate sim-

ulator for geographic routing in vehicular networks”, VANET’05