Mercury

Mercury: A Geo-Aware Mobility-Centric Messaging System
Teja Kommineni
teja.kommineni@utah.edu
Kirk Webb
kwebb@cs.utah.edu
School of Computing
University of Utah
50 S. Central Campus Drive
Salt Lake City, UT - 84112
ABSTRACT
Intelligent Transportation Systems require a communication
conduit for the agents in the ecosystem. A variety of safety-
and consumer-service messages are useful, even critical for
coordinating the ongoing state of mobile endpoints. Com-
municating telemetry between these endpoints, along with
data from other devices in the environment, allows them to
work together and improve safety. Moreover, it is important
for centralized services to have a global view and provide
information based on this vantage point. To this end, we have
developed Mercury, an end-to-end messaging system with
lightweight state overhead and publish/subscribe functional-
ity. Mercury is designed to flexibly integrate into the mobile
network environment under which it is deployed. Mercury’s
centralized Broker allows for data aggregation services and
data analysis. The Adapter tunes to the particular network
environment, allowing Client Endpoints a pathway to com-
municate. Concurrent processing and event-driven design
are used to minimize latency to meet demanding low-latency
messaging requirements.
1. INTRODUCTION
Intelligent Transportation Systems (ITS) [31] is an emerg-
ing area that includes many ideas and mechanisms for im-
proving the safety, quality of experience, and communication
capabilities of commuters. One particularly important as-
pect of ITS is vehicle-to-vehicle and infrastructure-to-vehicle
communication. Vehicle area networks (VANETs) [16] have
been introduced to meet the challenges of mobile network
actors with rapidly changing associations arising from dy-
namic proximity to infrastructure and other vehicles. To-
gether with mobile networking a la the 3GPP evolved packet
system [13], robust communication approaches involving
hybrid LTE and 802.11p [17] with dynamic multi-hop cluster-
ing have been proposed [28, 30, 32]. Considering safety and
other types of information exchange in an ITS, applications
and events have been identified that should be supported [11,
25], along with constraints such as latency for delivery and
scope (radius). Publish-subscribe systems provide a means
to efficiently distribute messages and events between infras-
tructure and mobile actors in an ITS. Such pubsub systems
can support peer-to-peer and infrastructure-sourced messages
at scale [24]. While research has been done in the areas
of VANET communication and mobility-aware publish/sub-
scribe systems, holistic compositions of these technologies is
lacking.
Transportation safety issues are also considerable. The
total number of vehicle sales in the United States averaged
15.43 million from 1993 up through 2015. The year 2015
saw a surge in sales, which climbed to 17.76 million vehicles
and is expected to rise with the improving economy [2]. In-
creasing vehicle populations will likely result in increasing
accident rates. A total of 32,675 people died in motor vehi-
cle crashes in 2014 [1]. We argue that technology advances
should be used to mitigate increasing casualties and ensure
safe journey. A key focus in this paper is using state-of-art
5G [3] networks to capture vehicle telemetry and pass along
safety events to vehicles that are relevant to their surround-
ings.
The setting in which any transportation messaging system
operates requires reliable, fast paced communication. How-
ever, the latency incurred by present solutions is not adequate
for time-sensitive interactions (e.g., for real-time inter-vehicle
notifications). Although there are architectures [28] that use
the 802.11p communication paradigm for low latency, they
lack good trust mechanisms. Such ad-hoc networks are also
often unstable, requiring frequent communication path adjust-
ments that result in losses and overhead. 3GPP-style mobile
networks provide advantages in stability, reachability, and
security.
This paper introduces Mercury; a integration of pubsub
systems and client endpoint mechanisms for low-latency mes-
sage transport. Part of the holistic vision that Mercury es-
pouses is mobility-specific geographic areas of interest (AOI).
These areas map to slices of physical locality that are rele-
vant to particular events. For example, a traffic accident and
resulting congestion are relevant to vehicles en route to the
accident location, back past potential egresses to alternate
routes. Mercury takes into account location-specific context
to calculate the relevant dynamic set of vehicles for message
transmission.
This paper makes the following contributions:
1

• A holistic end-to-end messaging system design.
The design of Mercury highlights the mechanisms by
which publish/subscribe systems and mobile endpoints
can be realized in 4G or 5G mobile network systems.
Broker aggregation services communicate through Adapters
to Client Endpoints at the edge. Mercury’s design ac-
commodates flexible deployment scenarios.
• A prototype implementation and evaluation of the Mer-
cury ITS messaging system.
We implemented a prototype of the Mercury compo-
nents and evaluated it on the PhantomNet [5] testbed.
This prototype makes use of OpenEPC [12] version 5,
including its simplified emulated radio access network
(RAN) environmnet. We drive the evaluation of Mer-
cury with our own mobility simulator, and report on the
former’s functional and latency characteristics.
The remainder of this paper is organized as follows: Sec-
tion 2 covers background, including putting Mercury into
context within related work, introducing mobile networking
environment concepts, and describing the vehicular endpo-
ing context. Section 3 provides an overview of the design
goals and components of Mercury, while section 4 delves
into the details of the design. Section 5 looks at how Mer-
cury can be positioned within different mobile networking
deployments. Section 6 briefly discusses the Mercury pro-
totype implementation. We cover our vehicle simulator and
provide an evaluation of the prototype in section 7. Section 8
discusses observations resulting from the design and imple-
mentation of Mercury, and where we think it should go next.
Finally, section 9 concludes.
2. BACKGROUND
The mobile network ecosystem is complex, and host to
a number of related efforts in providing context-specific
messages between endpoints in an Intelligent Transporta-
tion System environment. Before we discuss the design and
implementation of Mercury, we couch it in the context of
prior work. In addition, we provide backround on the 3GPP
4G Evolved Packet System [13], which is an important and
widely used mobile networking ecosystem that we position
Mercury within.
2.1 Related Work
There has been plenty of attention paid to effecient and
reliable delivery of messages within VANETs. Much of this
focuses on multi-hop clustering and hybrid use of evolved
packet system RAN (LTE). The VMaSC [28], MDMAC [30],
and NHop [32] systems attempt to form stable mobile 802.11p
clusters, using the LTE network to bridge between discon-
nected clusters. These solutions largely ignore the details of
the mobile core network, glossing over questions of compo-
nent placement and the resulting effects on latency. Moreover,
these systems are complimentary to Mercury in that they can
be used to reduce LTE resource contention and improve reli-
able transfer of messages. Such integration would however
come at the cost of additional complexity, failure modes, and
latency due to additional network path segments.
From the publish-subscribe perspective, there is a large
volume of prior work on traditional pubsub mechanisms [16]
[22] [19] [4]. This work is complementary to ours since it fo-
cuses on the useful aspects of pubsub, which we largely wish
to reuse. There has also been work on pubsub systems fo-
cused on mobile endpoints. The MoPS [24] publish-subscribe
system scales efficiently for large numbers of clients and
deals well with changing broker association. Mercury could
replace the Kafkasystem used in the prototype with MoPS
to better target mobile endpoints. However, MoPS does not
include the area of interest concept, which would continue
to be handled by the Mercury broker. A paper by Pongth-
awornkamol et al [26] looks at pubsub in the context of ad
hoc wireless networks. However, this work only performs
simulations of mechanisms, and does not consider a larger
top-down vantage point (important for coordination in a large
ITS).
Location-aware messaging, or geo-routing, in vehicular
networks has been studied fairly extensively [6]. Neverthe-
less, we find that most work has only proposed and simu-
lated mechanisms. Furthermore, many studies look at ad hoc
networks and fine-grained positioning of endpoints within
vehicle clusters. While such mechanisms may be helpful for
real-time collision avoidance, we argue that they tend to be
overly complex and unnecessarily constrain the communica-
tion domain to clusters of endpoints versus a central system
with a global view. The argument against centralized systems
is frequently that mobile networks are overloaded. This may
be true in instances of particularly high device concentration
(e.g. music festivals), but we found no studies showing a gen-
eral lack of available RAN resources outside of such crowded
contexts. Furthermore, in an ITS environment, vehicle popu-
lations and anticipated growth could be used to properly size
capacity.
The related work outline above focuses on particular as-
pects of messaging. To the best of our knowledge, no mes-
saging system targeted at mobile endpoints simultaneously
incorporates aspects we consider crucial to a holistic messag-
ing platform for Intelligent Transportation Systems. Such a
system should enable a global view of endpoints to facilitate
coordinated decision-making with complete data. It should
take advantage of mobile network environment mechanisms
(e.g. eMBMS [21]) to reduce overhead and latency. It should
consider the placement of components within the mobile
network and the impact of this placement. Finally, an ITS
messaging system should provide a location-aware address-
ing mechanism. The Mercury messaging system is designed
with all of these aspects in mind.
2.2 Mobile Networking Ecosystem
Mercury is primarily framed in the context of the 3GPP
2

4G and emerging 5G mobile networking architectures. There-
fore, we provide some background on these systems. The vast
majority of mobile carriers utilize these Evolved Packet Sys-
tems (EPS), making them an especially relevant environment
in which to operate a mobile messaging system. Note that
Mercury is also amenable to other mobility-friendly network
architectures, such as MobilityFirst [27], but we focus the
discussion in this paper on the 3GPP EPS. The 4G system has
been in active deployment since 2008 [9], and has undergone
a number of revisions. Also in play in some deployment
scenarios (see section 5) are software defined infrastructure
concepts that are expected to be prominent components in
the upcoming 5G EPS [3].
The 4G EPS includes the following key service functions
relevant to Mercury: Mobility Managment Entity (MME),
Home Subscriber Service (HSS), Serving Gateway (SGW),
Packet Data Network Gateway (PDN-GW or more commonly
PGW), evolved NodeB (eNodeB), and User Equipment (UE).
We will briefly describe the role of each of these components,
and their relationships with one another and with Mercury.
The Mercury architecture diagram in figure 1 shows Mercury
components in the context of a 4G EPS.
We will briefly cover the 4G components next. User
Equipment (UE) typically refers to end user devices such as
mobile phones, tablets, and 4G radio equipped laptops. The
Mercury Endpoint component runs on these. The Mobility
Management Entity (MME) is the 4G control plane func-
tion responsible for tracking the live (dynamic) session state
for UEs. The Home Subscriber Service (HSS) is essentially
a database of user (subscriber) information. The Serving
Gateway (SGW) is the first data tunnel anchor point that UE
sessions connect through (GTP tunnels). The Packet Data
Network Gateway (PGW) act as the egress point for a large
number of UE data bearers (GTP tunnels); they fan out to
multiple SGWs. Evolved NodeB (eNodeB) devices are the
wireless access points of the 4G EPS. They bridge the radio
access network (RAN) through which the mobile endpoints
(UEs) directly communicate with the evolved packet core
(EPC). GTP tunnels are established for each UE between
the eNodeB it is associated with and an upstream SGW. The
eNodeB also initiates session setup and default data bearer
establishment when UEs attach, acting as a proxy for UE to
MME control plane signalling. eNodeBs covering adjacent
cells coordinate through the MME and possibly with one
another to accomplish handover as endpoints move.
2.3 The Vehicle Environment
The environment of a vehicle endpoint is highly relevant
context for drivers and passengers. We argue that traffic,
safety issues, and proximity to resources are key to decision
making in this setting. Therefore, related telemetry such as
position, speed, and direction should be relayed for analy-
sis. Vehicle position is almost universally measured using
GPS [23]. Although known radio access points can be used
for rough localization, such positions are often too coarse-
grained. In fact, a better system would make use of computer
vision and/or static environmental sensors to extract lane po-
sition, relative distances to other vehicles, etc. Our work
uses GPS coordinates and leaves more fine-grained place-
ment techniques to future work. Speed is to be collected by
vehicle speedometers, or through monitoring of accelerom-
eters such as are found in most smart phones. Finally, to
acquire direction, digital compass readings can be gathered
from smart phones and are also available in many onboard
vehicle systems. We expect that environmental sensors will
grow in sophistication and accuracy through the proliferation
of self-driving cars and ITS.
3. DESIGN OVERVIEW
Delivering messages, important and casual, to mobile users
and endpoints that are interested in them is the overarching
premise of this work. Our vision for realizing an end-to-end
mobile message delivery service design principles revolve
around four central goals:
• Relevant content
The service should provide mechanisms to target groups
of endpoints which have explicit or implied interest in
messages. A message may be important because an
end user specifically asked for the content based on its
attributes (nearby gas prices). Alternatively, a message
may be deemed relevant for the endpoint because it is
related to an emergent event (vehicle accident ahead).
• Robust, low overhead, low latency communication
Message intent drives content delivery requirements.
For example, different types of emergency service mes-
sages have been identified for VANETs, each having
distinct latency requirements [11]. The service should
strive to minimize latency to provide on-time delivery
with headroom for outlier delays. Messages should
be categoriezed according to their relative importance
and processed accordingly (e.g., emergency info before
consumer content).
• Flexible deployment
Adoption of the service is bolstered by adaptability to
different mobile networking environments. Such ac-
comodation allows for deployment into LTE networks
with different geographic EPC service placements and
degrees of maleability. The service should allow for cen-
tralized metropolitan area integration (CloudRAN) and
distributed edge deployment (peer-to-peer eNodeB).
• Reuse of effective technologies
It is our contention that an end-to-end service should
not supplant existing mechanisms useful for achieve-
ing its composition. Indeed, it is counterproductive
to introduce new service components that induce un-
necessary changes and capital investments. The mes-
saging service should strive to work alongside existing
3

mobile network protocols and services. Only where
existing mechanisms do not provide key functionality
or do not give adequate service levels should changes
be introduced. Such changes should be as minimal and
transparent as possible to foster compatibility and ease
of adoption. On the other hand, considering less con-
strained future mobile network architectures [3, 29] is
also important.
We next describe the components of the Mercury messag-
ing system, along with how they fit into the 3GPP 4G mobile
networking architecture.
3.1 Mercury Components
Mercury is comprised of four essential components: mes-
sage broker, publish/subscribe system, message adapter, and
endpoints. These components are visible in the architecture
diagram in figure 1. This diagram shows the Mercury compo-
nents in the context of a 4G mobile network (the latter will
be described after Mercury is covered).
3.1.1 Message Broker
Mercury’s message broker is the brain center of the mes-
saging system. It processes all incoming messages and sends
these to relevant endpoints (via Kafka). The broker calculates
Areas of Interest (see section 4.2) for certain message types
(e.g., emergency notifications). Data analysis (aggregated
decisions, client statistics, etc.) occur at the broker. It does
not concern itself with how to get the messages to the target
endpoint(s); that job falls to the Mercury message adapter.
3.1.2 Publish/Subscribe System
Mercury needs a pubsub system that can serve a high
number of events with low latency. This is because of the
environment in which it operates. We have many agents/end-
points at any given point that interact with the system and for
them to effectively respond to a situation we need to dissipate
the messages quickly. The mobility of the clients creates the
potential for intermitent loss of connectivity with the pubsub
system which has to be re-established. Among the options
considered, Apache Kafka had the best combination of low-
latency and message queueing for recovering lost messages.
At a high-level Kafka gives the following guarantees: Mes-
sages sent by a producer to a particular topic partition will
be appended in the order they are sent. A consumer instance
sees records in the order they are stored in the log. For a topic
with replication factor N, kafka will tolerate up to N-1 server
failures without losing any records committed to the log.
3.1.3 Message Adapter
The Mercury message adapter is the conduit through which
pubsub messages flow through to endpoints. The adapter
coordinates sessions with client endpoints. It forwards client
reports and other content to and from the pubsub. It also
maintains pubsub topic subscriptions for the endpoints. The
prototype implementation described in this paper targets the
4G EPC, but it is possible to target other mobile networking
architectures with different adapters.
3.1.4 Endpoints
Mercury endpoints are the producers and consumers of
most content in the messaging system. Client Endpoints run
as applications on ITS-equipped vehicles. They report in with
telemetry, such as their current position and speed. Client
endpoints also subscribe to particular consumer messages
identified by specified attributes.
Service Endpoints may also be centralized services that
connect to the pubsub, running at the same location as the
Broker, and using the Broker’s services. Examples include
emergency notification processors, and consumer information
aggregators (gas prices).
3.2 Mercury Messages
Messages are the principle and only communication mech-
anism in Mercury. There are two high-level flavors: con-
trol and content. Control messages include session handling
(between broker and endpoints), broker to message adapter
communication, and subscription management (endpoint to
pubsub/broker). Content messages are those relevant to ap-
plications running on endpoints, such as emergency alerts
and consumer information (e.g. gas prices). Periodic session
maintenance messages are sent between the mobile endpoints
and the Mercury Adapter to update status (location) and track
liveness.
Content messages use structured message attributes to sig-
nal category information. This allows content producer and
consumer endpoints to steer messages to one another based
on interests. Most messages (including client reports) flow
through the pubsub system to the broker. This allows the
Broker to perform data aggregation and analysis (triggers
based on thresholds, for example).
Message destination addressing in Mercury includes end-
point, content, and area of interest targetting. Endpoint ad-
dresses are primarily used for session setup and teardown.
Content-specific addressing makes use of subscription at-
tributes to deliver messages. AOI addresses are unique to the
mobile environment. AOI bounds are computed by the broker.
These bounds form the destination address. Message adapters
check bounds to determine if their downstream area cover-
age is relevant before passing along, and endpoints check
their position relative to the bounds. In this way, Mercury
allows messages to be “area multicasted.” As we discuss in
section 4, our prototype restricts itself to simple bounding
models (radius). Complex spatial reasoning for improving
relevancy is left for future work.
4. DESIGN DETAILS
Each Mercury component runs as a self-contained applica-
tion. The Broker and Adapter(s) communicate over unmodi-
fied Kafka. Client Endpoints communicate with the Adapter
over UDP. Client-side applications using Mercury communi-
4

Figure 1: Mercury architecture diagram (in 4G mobile network context)
cate over the local loopback with the Client Endpoint process.
Service Endpoints use Kafkato communicate with the Broker
and talk to Clients via Adapter(s).
4.1 The messages and events
The Mercury Broker we have implemented handles differ-
ent types of events. Emergency Event, This is an event that
is published by a vehicle when it senses an emergency situa-
tion. Any message indicating such an event is immediately
processed by the Mercury broker and sends back an area of
interest (AOI) that has to be alerted for this event. Moving
Objects Event, This event indicates that there is an object that
is crossing a road. For such events, whenever a predicate cal-
culated by scheduler evaluates to true we inform the vehicles
in the AOI calculated. There are other types of events such
as Collision, Obstacle, Congestion and Blocked. All, these
events are handled similarly but they differ in two things. One
is the predicate function and the second is the frequency at
which the schedulers run. Our system also supports an other
type of event called Area Of Interest. This is different from
the aoi which is calculated by Mercury broker. In this type
of event the user is interested in knowing about the different
events happening at a particular location. Whenever an event
of type area Of interest is received by the Mercury broker. It
interacts with different handlers and determines the bin for
each event type in which the requested aoi falls into. Then
runs a predicate on these bins and learns about the traffic
conditions in that area which is communicated back to the
user.
Messages must fit into a single UDP packet (maximum of
64K). Each is self contained, with required identifiers, ad-
dresses, and message payload. There are two top-level types
of messages in the Mercury system: session and pubsub.
Session messages are transmitted between Client Endpoints
and associated Adapters. Pubsub messages can essentially
be exchanged between any Mercury components. Table 4.1
shows the fields present in Mercury Session and PubSub
messages. Each Mercury Address contains a destination and
source address. The source address can be an Adapter, a spe-
cific Client, a Service, or the Broker. The destination address
can be any of those listed for source, and additionally may
be a broadcast (all clients, or Area of Interest geo-address).
Each message contains key/value pairs in the payload which
are interpreted by specific applications. For example, emer-
gency alerts published by the Broker may be consumed by
an alert display application at the Client Endpoint. Client
Report message contain mobile telementry information: GPS
location, direction, and speed. These reports have semantic
meaning to Mercury, and so are tracked by the Client End-
point, Adapter, and Broker components. In our prototype
implementation, the Broker has had emergency event mes-
sage handling folded into it’s logic. Therefore, the prototype
Broker also interprets safety messages sent via pubsub top-
ics such as “Collision” and “Object_Hazard”. We envision
5

Field Description
UUID Unique identifier for message.
Session ID Identifies session between Client and Adapter.
Message Type Distinguishes intent (to/from Broker, PubSub, etc.)
Source Address Client ID, Service ID, Broker, or Adapter.
Dest Address Client ID, Service ID, Broker, Adapter, or AOI.
Topic Indicates topic message goes to or came from (PubSub only).
KeyVal Generic key-value pairs. Telemetry and other data is encoded here.
Table 1: Message fields found inside individual Mercury messages.
that such message handling would ultimately be moved to an
external Service Endpoint application.
4.2 Areas of Interest
Mercury uses Area of Interest geo-addressing to target
specific sets of Client Endpoints. When the Broker wishes to
send a message that is specific to a location context (e.g., a
warning triggered from multiple collision reports), it embeds
the desired geo-address as the destination, and sends this
along with the message toward the Clients. This message
is picked up from the pubsub by the Adapters. These check
their coverage area and forward the message along to Clients
within the defined geo-address that they serve (or drop if
there is no overlap). When an AOI-addressed message is
received by a Client Endpoint, it checks its current location
to determine if the message applies, and processes if so.
4.3 Mercury broker
Mercury Broker handles events in two steps: In the first
step events published are passed on to the respective han-
dlers where we have a filter that determines the different
locations from which the feeds are coming in for that specific
event.This will give us the messages published at different
locations for an event.Each location can be considered as a
bin. We maintain for each bin the count of messages pub-
lished, the center point and the radius for that bin. Whenever
a message comes into a bin it carries along the coordinates
from which this event has been published. We apply a simple
mean with the existing center point to determine the new
center point. Thus, at every instance we keep updating the
center point and radius.In the second step we have schedulers
for each event that get triggered at equal intervals based on
the event type. These schedulers empty the bins of their re-
spective event and run a predicate on them which is a simple
Boolean formula that checks for a condition. A predicate is
usually defined over some aggregation function expressed on
messages; when the predicate evaluates to true, the Mercury
broker publishes an event to the system with the center point
and the radius which we have calculated for that bin. These
are used to determine the area of interest(AOI), AOI is the re-
gion that is determined by the Mercury broker and constitutes
all the clients that have to be alerted about a situation.
4.4 Pubsub service
Within Mercury we require a publish-subscribe system for
communication between different entities. This requirement
directly arises from the need for each vehicle to send and lis-
ten to particular types of messages. Pubsub systems are well-
suited for these activities as we can publish and subscribe to
different events. In our system components such as Client
Endpoints and the Message Broker publish and subscribe to
events. Given this communication pattern, we developed the
set of requirments we need in a publish-subscribe mechanism
and surveyed available options.
As mentioned in section 3, we selected Apache Kafka as
our pubsub component. It is a distributed streaming platform
which lets users to publish and subscribe to streams of records.
Kafka is run as a cluster on one or more servers. The Kafka
cluster stores streams of records in categories called topics.
Each record consists of a key, a value, and a timestamp.
A topic is a category or feed name to which records are
published. In our case it is the event type. Topics in Kafka
can be multi-subscribed; that is, a topic can have one or more
consumers that subscribe to the data written to it. In our
implementation all the values written to a topic are listened
by Message Broker. But, we could extend this to different
service end points listening on the same topic.
For each topic, the Kafka cluster maintains a partitioned
log. Each partition is an ordered, immutable sequence of
records that is continually appended to a structured commit
log. The records in the partitions are each assigned a sequen-
tial id number called the offset that uniquely identifies each
record within the partition. In our system we create a single
partition for each topic. As within a partition order is main-
tained we can traverse through records using the offset of
the record. Each partition is replicated across a configurable
number of servers for fault tolerance. The Kafka cluster re-
tains all published records whether or not they have been
consumedâ ˘AˇTusing a configurable retention period.
4.5 Mobile network messaging adapter
The Mercury Adapter follows a multi-threaded, event-
driven execution pattern. It is highly modular, with clear
separation between client state tracking, network-specific ad-
dress mapping, and transport functionality. This allows new
modes of core network communication to be added alongside
(or in place of) existing mechanisms. Different threads of ex-
ecution handle pubsub, UDP communication, and scheduled
6

tasks (e.g., session checking); all are coordinated through the
main thread.
Mercury is designed so that any number of Adapters can
connect to the Kafkasystem. Each will send client telemetry
reports and pubsub messages along. A single broker message
sent via pubsub reaches all connected Adapters, which filter
as necessary. For example, an Adapter will drop messages
with Area of Interest destinations that do not overlap with its
coverage area. These coverage areas are configured by the
operator.
Endpoint identification/tracking The Adapter implements
client session tracking. When a client comes online, it goes
through a lightweight handshake with the Adapter. An INIT
message is sent to the adapter, which immediately responds
with an ACKNOWLEDGE message. The client includes a
telemetry report in its INIT message so that the Adapter and
Broker have immediate knowledge of its current state. Each
client is provided with a session ID token that must be used
in subsequent messages. The Adapter sends out periodic
HEARTBEAT messages to clients so that they can measure
liveness in the absence of other messages.
The Adapter maintains a small amount of state for each
client. This includes its unique ID, current mobile telemetry
(no history), session ID, and simple statistics (timestamp for
last message received, message count). The Adapter also
maintains a mapping from the client’s ID to its mobile core
network address. Changes in this address are automatically
detected and the mapping updated. The Adapter treats pe-
riodic client reports as heartbeats, and resets corresponding
individual timers as these are received. Sessions are pruned
after not receiving a report for three times the reporting inter-
val, or after receiving an explicit CLOSE message from the
client.
Unicast UDP transport In the prototype implementation,
we use UDP unicast packets for all communication between
Adapter and Client Endpoints. Ideally we would use eMBMS
broadcasts for AOI and broadcast (e.g. heartbeat) messages,
and we plan to do so in future work. Each UDP packet is a
self-contained Mercury message. The contents are encoded
using Google’s Protocol Buffers [7] for performance and
space efficiency. For AOI destination addresses, the Adapter
calculates the set of clients to transmit a message to and
unicasts to each individually. For broadcasts, it transmits to
all clients with valid sessions.
Pubsub passthrough Messages from established clients
that are intended for the pubsub system are directly published
to the indicated topic by the Adapter. Likewise, messages
received from the Broker via pubsub are sent to specific
endpoints or broadcast to all clients handled by the Adapter
(as appropriate).
4.6 Client Endpoint mechanisms
The Mercury Client Endpoint connects to the Adapter
over UDP. It establishes a session using a simple handshake;
an INIT message is sent to the adapter, which it replies to
with an acknowledgement. The client will continue trying to
connect until it receives the acknowledgement, backing off
linearly with random jitter to prevent synchronization with
other clients (prevent in-cast overload). The client will also
try to reestablish its session if it does not receive a heartbeat
from the Adapter after three heartbeat intervals (tunable).
Messages are encoded using Google’s Protocol Buffers.
The client gathers movement telemetry via the local sen-
sors of the device it is running on and sends these along pe-
riodically as report messages to the Adapter. As mentioned,
the Adapter uses these reports to measure the liveness of the
connected clients. Some random jitter is added initially to
the report interval to again avoid synchronization with other
clients.
The client listens over the loopback interface for connec-
tions from local applications. Similar to what the Adapter
does for its Client Endpoints, a Client Endpoint relays sub-
scription requests and messages from applications through to
the pubsub/broker. Protocol Buffers are also used between
clients and applications for encoding the messages.
5. DEPLOYMENT SCENARIOS
A key aspect of the Mercury design is adaptability to dif-
ferent mobile network deployments and architectures. End-
points, broker, and pubsub components are all mobile-environment
agnostic. Mobile endpoint identity is specific to the Mercury
messaging system. One of the message adapter’s primary
functions is to separate the concerns of the messaging sys-
tem with the routing of messages inside the mobile network.
Mercury requires a low-latency vantage point near the mobile
network edge. Message adapters, pubsub, and broker all need
to reside at this vantage point. Such proximity allows for
message delivery to meet timing requirements. Distance to
endpoints incurs a tradeoff between convenience of deploy-
ment and ability to meet timing needs. We show deployment
versus distance/latency requirements in table 5.
ADD REFERENCES TO TABLE
Latency Tolerance Examples Ref
High (> 1 second) Consumer info, Congestion [11, 25]
Medium (< 50 msec) Lane change assistance [18]
Low (< 10 msec) Vehicle control loop [15]
Table 2: Latenct tolerances for different types of vehicu-
lar environment messages.
The message adapter can utilize different techniques to
interface with a mobile network deployment. In the sim-
plest case, it can interact with endpoints via routed network
addresses (e.g. IPv4 or IPv6). Alternatively, it may trans-
parently interpose on encapsulated communication channels
to inject/extract messages (e.g. GTP bearers). Another op-
tion is for the message adapter to operate in concert with
the wireless access point right at the edge. It could do this
by transparently interposing on the access point’s data plane
7

Figure 2: Mercury deployment scenarios.
uplink, or via an integrated side channel added into the access
point software. We next present three plausible deployment
scenarios, each of which is shown in figure 2.
5.1 Metropolitan area 4G deployment
Consider a city investing in Intelligent Transportation Sys-
tem infrastructure. Resources include a data center within the
city and wireless coverage via eNodeBs (particularly along
major vehicle corridors). The 4G EPC components reside
in the local data center, including the PGW. Given a maxi-
mum cabling distance from the data center to any eNodeB
of around 100 kilometers, one way speed of light latency
is bounded to less than 5 milliseconds. LTE one-way wire-
less first-hop target latency is 5 ms, but may be more like
10 ms [20] due to UE and eNodeB processing overhead. Fur-
ther switch/router hops within the data center should not
appreciably add to the latency. Egress through SGW/PGW
can add another 5 ms of one-way delay. Total mobile net-
work infrastructure RTT is thus bounded to 40 ms. This
leaves 10 ms for Mercury components to process messages
for low-latency vehicle-to-vehicle coordination (within 50
ms).
In this deployment, the Mercury message adapter, pubsub,
and broker all reside alongside the PGW in the local metro
data center. The message adapter and mobile endpoints com-
municate using native network addresses (i.e. IPv4 or IPv6).
Mercury components communicate via the data center pub-
sub deployment. The endpoints can obtain the address of
their associated message adapter using DNS, DHCP options,
or multicast query.
5.2 Existing mobile carrier 4G network deploy-
ment
In an existing 4G mobile carrier network, there may not be
a centralized upstream network egress location in a particular
metro area of interest. Further, PGW nodes are typically
deployed at a handful of national data centers. Total one-way
latency to egress through a PGW in such deployments is
often over 50 ms [14, Chapter 7]. Installing Mercury at these
national data centers would impair its ability to meet SLAs
for some classes of low-latency messages (see table 5). Many
mobile carriers, however, concentrate regional connectivity at
a Mobile Telephone Switching Office (MTSO). Typical one-
way latency from eNodeB to MTSO is 10 ms [10]. With LTE
RAN latency added, the one way latency between MTSO and
UE is about 20 ms.
Deployed into a MTSO, Mercury message adapters can
interpose on GTP data bearers for UEs in the same way
that SMORE [10] does. Alternatively, lower-capacity (thus
cheaper) combined SGW/PGW functions may be deployed
into the MTSO. This would allow the provider to setup dedi-
cated bearers for services such as Mercury in a MTSO loca-
tion. Using these options, Mercury can operate with slightly
higher RTT compared to metropolitan data center deploy-
ments. However, the additional latency may not meet the
latency requirements for some UE-to-UE (vehicle-to-vehicle)
applications.
5.3 Mobile edge 4G/5G deployments
Mercury can be altered to run at the network edge, on the
mobile wireless access points. If a deployment includes the
ability for running services on the eNodeB (or equivalent)
access points, then Mercury can operate using peer-to-peer
semantics. All components of Mercury would run on each
eNodeB, and some service endpoints as well. The pubsub
is extended to form a connectivity mesh between adjacent
eNodeBs. Mercury’s scope at each eNodeB is smaller, and
so the amount of processing is reduced. When a message
8

destination is an area of interest, the local Mercury broker
determines whether parts of this area lie outside of its cover-
age. If so, it sends such messages along (via pubsub mesh)
to neighboring eNodeBs. These handle their portion of the
endpoints in the AOI. Arbitrary consumer content messages
are also sent along to potential subscribers via the pubsub
mesh. This deployment scenario allows for very low-latency
messaging between nodes connected via the same eNodeB
(potentially under 20 ms RTT considering RAN latency).
Inter-eNodeB communication latency depends on the com-
munication path between eNodeBs. This could be only a
handful of milliseconds if there are metro-area-local paths.
A second realization of this deployment scenario is via
a 5G CloudRAN environment [8]. Here eNodeBs are split
into remote radio heads (RRH) connected to base band units
(BBUs). The latter are located in nearby compute aggrega-
tion locations; RRH and BBU functions can only be about
25 km apart in order to maintain the closed control loop
between them. BBU functionality in a compute location
serves multiple RRH units. Mercury components would be
deployed into the compute locations in this case. Multiple
such compute locations may serve an area. Mercury could
be deployed to each with a peer-to-peer pubsub mesh setup
between them. Similar very low latency messaging is possi-
ble in this scenario with a Mercury message adapter working
with the BBUs at its location. An additional advantage is that
the very low latency messaging extends to the (larger) area
covered by all associated RRH units.
6. IMPLEMENTATION
We next briefly cover some of the essential implementa-
tion details of the Mercury prototype. Section 4 discusses
our selection of Kafka for the pubsub component. Google’s
Protocol Buffers were used for efficient on-the-wire binary
encoding of messages. Such encoding is used on all non-
pubsub communication paths. The Mercury prototype was
written in python in about 2,200 lines of code. We chose
Python because it was easy to rapidly produce a working
system in this high-level langauge, and because it has good
support for parallel processing. Integrations exist for Python
for the other technologies used. Concurrent threads are em-
ployed to capture and buffer incoming messages from both
the pubsub and UDP point-to-point channels. An internal
scheduler was implemented to perform periodic tasks, and
supports randomized offsets.
The prototype implementation targets the centralized PGW
deployement scenario where the core Mercury components
(Adapter, Broker, and pubsub) are positioned immediately
on the other side of a standard 3GPP data bearer egress
point. This deployement scenario affords low latency over-
head when EPC core components are located in a datacenter
near the network edge. We covered this deployement scenario
in section 5.1. Note that we did not take advantage of any
specific aspects of the mobile environment in our prototype,
which is left as future work.
7. EVALUATION
We have performed functional and performance (latency)
evaluations of our Mercury prototype. The components oper-
ate functionally as expected, tracking location and responding
to area of interest events where they are verified to be in the
defined region. Micro-benchmarks show that the system can
operate down to 10ms RTT from baseline single message
testing, with a slowly growing response that begins to ramp
up more significantly once the input message rate overtakes
the maximum processing rate of the components. We believe
that introducing more concurrency will both lower the nomi-
nal RTT even for moderate incoming message rates, and push
out the maximum processing rate.
7.1 Vehicle Mobility Simulator
To verify Mercury, we have written a simulator that mimics
real-time vehicular movement. The Simulator provides two
main functionalities. Firstly, to add vehicles to the system.
Secondly, to emulate an event at a given location. Mercury
Simulator has been implemented as follows: We define an
area with in the co-ordinate system in which the vehicles
keep moving. This area depends upon the number of vehicles
a user wants to initialize. Each vehicle is associated with a
unique identifier, x and y co-ordinates that indicate its loca-
tion in the system and speed at which it is moving. Each
vehicle on being initialized binds themselves to a port and
periodically sends reports to the Message Adapter about its
location. While sending a report, we take the vehicles present
speed and its location to determine its new location. Which
in turn is updated on the vehicle and communicated to the
Message Adapter. This is done for each vehicle and thus we
have a simulated vehicle environment. One point to be noted
here is since it is a bounded region we take care that all the
vehicles move within this enclosed area.
Using simulator, we can emulate an event by specifying the
(x,y,r) 3-tuple. All the vehicles enclosed within the region
covered by a circle centered at (x,y) with radius r sense this
event. We have implemented this as follows: Whenever a
simulate event function is invoked with 3-tuple. Since we
have list of all the vehicles within the system for each vehicle
we verify if it falls within the region specified by the 3-tuple.
If yes, we send a message to the Message Adapter on the port
that vehicle is bound.
7.2 Functional Evaluation - Area of Interest
Using the Message Simulator we have verified the func-
tionality of the system. To do this we have initially set up ten
vehicles in the simulated ecosystem. Then mocked an event
collision at location (5,2) with a radius of 2 as in figure 3.
All the vehicles within this region have sensed the event and
published a message to the Mercury. We have taken care
that the number of vehicles that sense the event and publish
the messages to Message Adapter will be greater than the
threshold needed for the Message Broker to trigger an Area
of Interest message. The AOI message from Message Broker
9

is represented as bigger circle in the figure 4.
Figure 3: Mercury event simulation.
Figure 4: Mercury Area Of Interest.
By the time AOI message from Message Broker has reached
the vehicles we have observed the vehicle which has sensed
the event at location (6,2) has moved away from the region.
This vehicle didn’t receive any alert from Mercury. There
were two other vehicles that are at location (3,5) and (8,6)
and have moved into the AOI calculated by Message Broker
both these vehicles received the message. And the three of
four vehicles which have sensed the event were still falling
within the Message Broker’s AOI and have received an alert.
Through this experiment we have verified that all the vehi-
cles within the AOI of Message Broker have received an alert
irrespective of sensing the event and those vehicles which are
out of the AOI did not receive it.
7.3 Latency Micro-benchmark
In order to assess the latency response of Mercury under
load, we performed a micro-benchmark involving sending
messages at progressively higher rates, end-to-end. To do
this, we developed an client-side application, “echorate”,
that sends timestamped messages to the Broker. The Broker
responds to these immediately, embedding the timestamp
provided (similar to ICMP ping). The “echorate” app takes
rate and total message count as arguments. For this evaluation,
we took baseline measurements at 1 message per second over
a one minute period, then at rates of 10 messages/second up
through 1,000 messages/second at roughly logarithmic rate
increments. For each increment we increase the count such
that the total test target run time is 10 seconds.
Figure 5: Round trip time (RTT) as a function of message
rate.
The results are presented as a box-and-whisker plot in
figure 5. Message rate appears as discrete values along the
x-axis. Round trip time is plotted using a log scale on the
y-axis. The dots indicate outliers. This graph shows that
the minimum RTT our prototype manages is about 10 mil-
liseconds. However, even at a rate of 10 messages/second,
the RTT jumps by an order of magnitude. The response
stays fairly flat at around 1 second through a rate of 500
messages/second. Somewhere between this latter rate and
750 messages/second, the rate begins a linear climb. This
trend is a result of the Mercury components falling behind in
message processing. As the quartile spread indicates at these
higher rates, the RTT slowly creeps up as message processing
continues to lag behind the incoming message rate.
These results indicate that the prototype needs additional
work to improve its throughput and responsiveness in order
to meet the demand of realistic workloads while maintaining
latency-sensitive message requirements. Our prototype al-
10

ready uses parallel processing inside individual components
to queue incoming messages. However, message processing
and sending is single-threaded. We believe that the key to
increasing throughput is to parallelize these tasks. Decreas-
ing the RTT baseline message rate can be accomplished by
reimplementing Mercury in a lower-level language where
individual processing time contributions are easier to manage
and processing overhead is generally lower. Using real-time
scheduler directives would likely also help. We leave such
optimizations as future work.
8. DISCUSSION
The design of Mercury, the possible deployment scenarios
it can accommodate and the evaluation of prototype imple-
mentation of Mercury clearly suggest that Mercury is a fit for
both present and future Intelligent Transportation Systems.
Yet, there are few areas where we can further improve to
make Mercury more appealing.
At present, Mercury uses AOI as a main primitive in deliv-
ering messages to the clients. Mercury Broker determines the
AOI and this is processed by the Message Adapter to deliver
messages to respective Clients. The AOI that we are imple-
menting or envisioned in this paper is based on the circle
centered at a point with a given radius. Clearly, when we are
using the circular region as metrics there could be vehicles
which are not affected by the event yet receive the alert mes-
sage from Mercury because they fall inside this region. This
is where we think AOI calculation should also add another
dimension to its measurement, namely direction. This comes
from the basic concept that an event happening on one side
of the road may or may not affect the vehicles moving on the
other side of the road. Looking at the directions we can also
detect collisions between vehicles using the reports they send
to the system. Hence, this is an interesting future work area.
Presently, the Message Adapter doesn’t implement any sig-
nificant security measures when dealing with the messages
that are moving in and out of the system. We only check
if the sender of the message has a session with Mercury
and then forward it to the Mercury Broker. The verifica-
tion that is done here can be further supplemented by other
methods such as self-certifying endpoint IDs and signed mes-
sages (to prevent spoofing), and data analysis at the Broker
looking for anomalies (to prevent gaming). Without such
protections, miscreants could exploit and/or spoof messages
without sensing an event or in any other anticipated way that
could generate profits to them or cause harm to others.
The Message Broker receives all sensed events from the
Message Adapter and performs computation on them to de-
termine AOI. Here, we can see that Message Broker depends
on the incoming messages to figure out whether an inciden-
t/event is happening. It doesn’t have any other way to know
about an event except for the incoming messages. But, if we
have a close look at Mercury we also have another source of
information; the reports that are periodically sent by vehicles.
We can make the Message Broker intelligent enough to look
at these reports and detect an event. It could further decrease
the latency in the system and we could make more informed
decisions a step before a threshold of vehicles get to sense
this event and report it to the system. For example, all vehi-
cles which are in a region experiencing congestion move at a
slow speed. Looking at the reports from vehicles, we could
detect the slowing down of vehicles in any region and send a
congestion alert to relevant vehicles (via AOI).
Section 5 mentions a handful of different deployment sce-
narios. We would like to extend Mercury into these different
scenarios and measure its performance. In particular, we
would like to make use of eMBMS for broadcasting AOI and
heartbeat messages to Clients. We would also like to explore
the edge deployment scenario. This scenario has the great-
est potential for minimizing latency, and also has interesting
fault tolerance properties and challenges. Keeping a mesh of
agents running at each eNodeB bypasses issues in the core
network, but means manging message routes and maintaining
healthy paths.
Finally, there are a number of things we can do to improve
the performance of Mercury. We could increase the perfor-
mance by increasing parallelism in the system such as scaling
Message Adapter and Message Broker. We could also take
advantage of the consumer groups concept in the Apache
Kafka that help us in instantiating multiple consumers on a
single topic. This way all the messages that come to a single
topic can also be processed by multiple consumer instances
giving us high throughput. We can also take advantage of the
retention period in the Kafka to go back in time and receive
any alert if missed. This could probably be an interesting op-
tion to look at because vehicles may tend to lose connection
to intermittently and when they establish connection again
with the system they might be interested in listening those
events that it missed.
9. CONCLUSION
We have presented Mercury, an end-to end mobility centric
messaging system that improves safety within the vehicular
environment. We have looked at the architecture and differ-
ent components of Mercury. We have discussed the different
deployment scenarios and how Mercury can fit in each of
them. We also analyzed the need of a pubsub system in this
environment and our choice to use Kafka for the implementa-
tion. We evaluated our prototype Mercury implementation in
a 4G deployment context. A vehicle simulator was written
to test functional correctness, and a synthetic echo applica-
tion was used to evaluate rount trip time latency performance.
We conclude that the prototype is functionally accurate, but
requires additional parallelism and tuning to meet latency re-
quirements for realistic workloads. We have also considered
how adding a direction based metric to this system would
make it more precise and how we can incorporate more trust
into Mercury so that it is not misused.
10. REFERENCES
11

[1] Insurance institute for highway safety. http:
//www.iihs.org/iihs/topics/t/general-statistics/
fatalityfacts/overview-of-fatality-facts, 2015.
[2] Trading economics. http://www.tradingeconomics.com/
united-states/total-vehicle-sales, 2015.
[3] 5GPPP. 5G Vision - The 5G Infrastructure Public Private Partnership:
the next generation of communication networks and services.
https://5g-ppp.eu/wp-content/uploads/2015/02/
5G-Vision-Brochure-v1.pdf, 2015.
[4] ARANITI, G., CAMPOLO, C., CONDOLUCI, M., IERA, A., AND
MOLINARO, A. Lte for vehicular networking: a survey. IEEE
Communications Magazine 51, 5 (2013), 148–157.
[5] BANERJEE, A., CHO, J., EIDE, E., DUERIG, J., NGUYEN, B.,
RICCI, R., VAN DER MERWE, J., WEBB, K., AND WONG, G.
Phantomnet: Research infrastructure for mobile networking, cloud
computing and software-defined networking. GetMobile: Mobile
Computing and Communications 19, 2 (2015), 28–33.
[6] BILAL, S. M., BERNARDOS, C. J., AND GUERRERO, C.
Position-based routing in vehicular networks: A survey. Journal of
Network and Computer Applications 36, 2 (2013), 685–697.
[7] BUFFERS, P. Googleâ ˘A´Zs data interchange format, 2011.
[8] CHECKO, A., CHRISTIANSEN, H. L., YAN, Y., SCOLARI, L.,
KARDARAS, G., BERGER, M. S., AND DITTMANN, L. Cloud ran for
mobile networksâ ˘AˇTa technology overview. IEEE Communications
surveys & tutorials 17, 1 (2015), 405–426.
[9] CHEN, Y., DUAN, L., AND ZHANG, Q. Financial analysis of 4g
network deployment. In 2015 IEEE Conference on Computer
Communications (INFOCOM) (2015), IEEE, pp. 1607–1615.
[10] CHO, J., NGUYEN, B., BANERJEE, A., RICCI, R., VAN DER
MERWE, J., AND WEBB, K. Smore: software-defined networking
mobile offloading architecture. In Proceedings of the 4th workshop on
All things cellular: operations, applications, & challenges (2014),
ACM, pp. 21–26.
[11] CONSORTIUM, C. V. S. C., ET AL. Vehicle safety communications
project: Task 3 final report: identify intelligent vehicle safety
applications enabled by dsrc. National Highway Traffic Safety
Administration, US Department of Transportation, Washington DC
(2005).
[12] CORICI, M., GOUVEIA, F., MAGEDANZ, T., AND VINGARZAN, D.
Openepc: A technical infrastructure for early prototyping of ngmn
testbeds. In International Conference on Testbeds and Research
Infrastructures (2010), Springer, pp. 166–175.
[13] DAHLMAN, E., PARKVALL, S., AND SKOLD, J. 4G:
LTE/LTE-Advanced for Mobile Broadband, 1st ed. Academic Press,
2011.
[14] GRIGORIK, I. High Performance Browser Networking: What every
web developer should know about networking and web performance. "
O’Reilly Media, Inc.", 2013.
[15] HANSSON, H. A., NOLTE, T., NORSTROM, C., AND PUNNEKKAT,
S. Integrating reliability and timing analysis of can-based systems.
IEEE Transactions on Industrial Electronics 49, 6 (2002), 1240–1250.
[16] HARTENSTEIN, H., AND LABERTEAUX, L. A tutorial survey on
vehicular ad hoc networks. IEEE Communications magazine 46, 6
(2008), 164–171.
[17] JIANG, D., AND DELGROSSI, L. Ieee 802.11 p: Towards an
international standard for wireless access in vehicular environments.
In Vehicular Technology Conference, 2008. VTC Spring 2008. IEEE
(2008), IEEE, pp. 2036–2040.
[18] KARAGIANNIS, G., ALTINTAS, O., EKICI, E., HEIJENK, G.,
JARUPAN, B., LIN, K., AND WEIL, T. Vehicular networking: A
survey and tutorial on requirements, architectures, challenges,
standards and solutions. IEEE Communications Surveys & Tutorials
13, 4 (2011), 584–616.
[19] KIM, H.-Y., KANG, D.-M., LEE, J.-H., AND CHUNG, T.-M. A
performance evaluation of cellular network suitability for vanet. World
Academy of Science, Engineering and Technology, International
Journal of Electrical, Computer, Energetic, Electronic and
Communication Engineering 6, 4 (2012), 448–451.
[20] LANER, M., SVOBODA, P., ROMIRER-MAIERHOFER, P., NIKAEIN,
N., RICCIATO, F., AND RUPP, M. A comparison between one-way
delays in operating hspa and lte networks. In Modeling and
Optimization in Mobile, Ad Hoc and Wireless Networks (WiOpt), 2012
10th International Symposium on (2012), IEEE, pp. 286–292.
[21] LECOMPTE, D., AND GABIN, F. Evolved multimedia
broadcast/multicast service (embms) in lte-advanced: overview and
rel-11 enhancements. IEEE Communications Magazine 50, 11 (2012),
68–74.
[22] MIR, Z. H., AND FILALI, F. Lte and ieee 802.11 p for vehicular
networking: a performance evaluation. EURASIP Journal on Wireless
Communications and Networking 2014, 1 (2014), 1.
[23] MISRA, P., AND ENGE, P. Global Positioning System: Signals,
Measurements and Performance Second Edition. Lincoln, MA:
Ganga-Jamuna Press, 2006.
[24] NASIM, R., KASSLER, A. J., ANTONIC, A., ET AL. Mobile
publish/subscribe system for intelligent transport systems over a cloud
environment. In Cloud and Autonomic Computing (ICCAC), 2014
International Conference on (2014), IEEE, pp. 187–195.
[25] PAPADIMITRATOS, P., DE LA FORTELLE, A., EVENSSEN, K.,
BRIGNOLO, R., AND COSENZA, S. Vehicular communication
systems: Enabling technologies, applications, and future outlook on
intelligent transportation. IEEE Communications Magazine 47, 11
(2009), 84–95.
[26] PONGTHAWORNKAMOL, T., NAHRSTEDT, K., AND WANG, G. The
analysis of publish/subscribe systems over mobile wireless ad hoc
networks. In Mobile and Ubiquitous Systems: Networking & Services,
2007. MobiQuitous 2007. Fourth Annual International Conference on
(2007), IEEE, pp. 1–8.
[27] RAYCHAUDHURI, D., NAGARAJA, K., AND VENKATARAMANI, A.
Mobilityfirst: a robust and trustworthy mobility-centric architecture
for the future internet. ACM SIGMOBILE Mobile Computing and
Communications Review 16, 3 (2012), 2–13.
[28] UCAR, S., ERGEN, S. C., AND OZKASAP, O. Multihop-cluster-based
ieee 802.11 p and lte hybrid architecture for vanet safety message
dissemination. IEEE Transactions on Vehicular Technology 65, 4
(2016), 2621–2636.
[29] VENKATARAMANI, A., KUROSE, J. F., RAYCHAUDHURI, D.,
NAGARAJA, K., MAO, M., AND BANERJEE, S. Mobilityfirst: a
mobility-centric and trustworthy internet architecture. ACM
SIGCOMM Computer Communication Review 44, 3 (2014), 74–80.
[30] WOLNY, G. Modified dmac clustering algorithm for vanets. In 2008
Third International Conference on Systems and Networks
Communications (2008), IEEE, pp. 268–273.
[31] ZHANG, J., WANG, F.-Y., WANG, K., LIN, W.-H., XU, X., AND
CHEN, C. Data-driven intelligent transportation systems: A survey.
IEEE Transactions on Intelligent Transportation Systems 12, 4 (2011),
1624–1639.
[32] ZHANG, Z., BOUKERCHE, A., AND PAZZI, R. A novel multi-hop
clustering scheme for vehicular ad-hoc networks. In Proceedings of
the 9th ACM international symposium on Mobility management and
wireless access (2011), ACM, pp. 19–26.
12

Mercury

More Related Content

What's hot (18)

Similar to Mercury (20)

Mercury