4726 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 23, NO.

5, MAY 2022

Graph-Enabled Intelligent Vehicular

Network Data Processing
Zhigao Zheng , Member, IEEE, and Ali Kashif Bashir , Senior Member, IEEE

Abstract— Intelligent vehicular network (IVN) is the under-

lying support for the connected vehicles and smart city, but
there are several challenges for IVN data processing due to the
dynamic structure of the vehicular network. Graph processing,
as one of the essential machine learning and big data processing
paradigm, which provide a set of big data processing scheme,
is well-designed to processing the connected data. In this paper,
we discussed the research challenges of IVN data processing
and motivated us to address these challenges by using graph
processing technologies. We explored the characteristics of the
widely used graph algorithms and graph processing frameworks
on GPU. Furthermore, we proposed several graph-based opti-
mization technologies for IVN data processing. The experimental
results show the graph processing technologies on GPU can
archive excellent performance on IVN data.
Fig. 1. An example structure of vehicular networks. V2D is the com-
Index Terms— Graph processing, vehicular networks, internet munication of vehicle-to-device, while the V2V is the communication of
of intelligent vehicles. vehicle-to-vehicle.

I. I NTRODUCTION
for current vehicular communication systems with higher

W IRELESS network is the basis of many applications

that can support mobile broadband access. In recent
years, the wireless network attracted more and more attention
security, reliability, and effectiveness requirements. Along with
recent advances in some new technologies such as computa-
tional intelligence, artificial intelligence (AI), blockchain, and
from both industry and academics [1], [2]. Vehicular net- 5G will help the ITS move forward into a smarter level with
work is a smart-vehicles network with complex intra-vehicle autonomous driving in near further.
systems [3]–[5]. Various vehicular infrastructures (such as In recent years, there are many academic and industry
cameras, traffic lights) and kinds of sensors, which used to organizations, and even some government departments also
collect vehicle and driving status, are included in the vehicular proposed a set of communication standards for vehicular
network. The general architecture of the vehicular network Ad Hoc networks (VANETs), including the Dedicated Short-
shown in Fig. 1, there are three layers (vehicles, connections, Range Communication (DSRC) [9], and ITS-G5 [10], [11] in
and servers/clouds) that included in this architecture. The vehi- the United States and Europe respectively, both of them are
cles communicate with each other through the V2V (vehicle proposed based on IEEE 802.11p [12]. However, some recent
to vehicle) network, at the same time, the vehicles can com- researches [13], [14] show the communication data processing
municate with remote mobile phones and some other devices technologies suffered from several issues, such as lack of
through V2I (vehicle to infrastructure) network. Promoted by quality of service (QoS) guarantees, channel access delay,
the new onboard computing and sensing technologies, the different communication standard transfer, and short-lived
vehicular networks have become one of the most important vehicle-to-infrastructure (V2I) connection. In order to solve all
aspect of the intelligent transportation systems (ITS) and smart these problems, the 3rd Generation Partnership Project (3GPP)
cities [6]–[8]. This new trend proposed some new challenges started several investigations to find the vehicle-to-everything
(V2X) supporting services for the long term evolution (LTE)
Manuscript received January 30, 2020; revised June 21, 2020; accepted
September 8, 2020. Date of publication April 5, 2022; date of current network and the future 5G cellular system [15]. With the devel-
version May 3, 2022. The Associate Editor for this article was S. Mumtaz. opment of device banded communications, some researchers
(Corresponding author: Zhigao Zheng.) try to employ the device-to-device (D2D) communications
Zhigao Zheng is with the School of Computer Science and Technology,
Huazhong University of Science and Technology, Wuhan 430074, China to support the vehicle-to-vehicle (V2V) transmission in cel-
(e-mail: zhengzhigao@[Link]). lular systems [6], [16], [17]. In addition, recent research
Ali Kashif Bashir is with the Department of Computing and Mathematics, have extended the graph algorithms and some graph enabled
Manchester Metropolitan University, Manchester M15 6BH, U.K. (e-mail:
[Link].b@[Link]). communication data processing framework for vehicular net-
Digital Object Identifier 10.1109/TITS.2022.3158045 works resource allocation [16]–[19]. However, there are still
1558-0016 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See [Link] for more information.

Authorized licensed use limited to: PSG COLLEGE OF TECHNOLOGY. Downloaded on July 06,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
ZHENG AND BASHIR: GRAPH-ENABLED INTELLIGENT VEHICULAR NETWORK DATA PROCESSING 4727

some challenges in designing graph enabled vehicular network of the graph processing approach for intelligent vehicular
communication data processing, such as how to provide QoS network data processing framework, Hooker. We also designed
guarantees for the heterogeneous network, and how to ensure a set of experiments to verify that the idea of applying the
the computation results in the dynamic vehicular environment. graph processing approach into intelligent vehicular network
Meanwhile, with the development of sensing technolo- communication data processing is extremely appealing for
gies and 5G, the further vehicles will be equipped with more several reasons.
and more smart sensors, such as the radar, engine monitor The rest of the paper is organized as follows. We intro-
and control unit, light control unit and some other device duce the characteristics and research challenges of intelligent
such as remote emergency alarm device as well as some vehicular network data processing and the motive of using the
networks based applications such as social networking appli- graph processing method to process the large scale intelligent
cations. Most of these sensors need to be work in real-time, vehicular network data in section II. We introduce the major
together with the high-performance computing and storage graph processing frameworks and algorithms in section III
unit onboard, and these sensing technologies will promote and IV. Then, we introduce how we can use graph algorithms
the intelligent driving progress into autonomous driving. All to solve the challenges of intelligent vehicular network data
these sensors and onboard devices keep collecting, generating, processing, and also the design methodologies and principles
storing, and communicating with each other, which generates of Hooker in section V. The experimental evaluation of the
large scale connected data. There are complex connections existed graph processing frameworks introduced in section VI.
between all these communicating and controlling data, how to At last, we conclude the paper in section VII.
explore the relations between these data to improve the control
and communication performance for all these devices is a great II. C HALLENGES OF IVN C OMMUNICATION
challenge. DATA P ROCESSING
The graph is a fundamental mathematical structure used High dimensional intelligent vehicular network data exhibit
to model pairwise relations between objects, and it is widely distinctive characteristics, which have created significant chal-
used in machine learning [20], [21], and deep learning [22] lenges to intelligent vehicular network design. In this section,
technologies to express the connections between different we discuss these challenges and try to figure out the potential
objects. The context in a graph is called vertices (also solutions by using graph processing technologies.
called nodes), while the connections are called edges. Graph
theory has been widely used in vehicle communications A. Highly Dynamic Topology
[23], [24], a graph-based metrics is proposed to gauge the
redundancy of dissemination protocols in [25], some open Unlike typical mobile network graphs, the vertices represent
issues of the metrics such as attackers colluding and eviction vehicles are moving at quite high speed, which leads the
are also included in this work. The authors of [26] proposed a topology of vehicular network changes frequently. This char-
VANET communication model based on evolving graph theory acteristic will affects system design and computing in multiple
to determines the regular routes preemptively. The authors aspects of vehicular networks. On the one hand, the network
of [27] formulated the problem of cooperative communications connection performance is much worse than a common mobile
scheduling in vehicular networks by using graph theory, and network, and the link failures and message loss rate will
the authors proposed a bipartite-graph-based (BG) scheduling also be increased, under this scenario. How to elongate the
scheme to allocate the vehicle-to-infrastructure (V2I) and life of communication connections and how to manage the
V2V links for both single-hop and dual-hop communications. connection links between the moving vehicles is an excellent
The experimental result shows that the proposed method can challenge for intelligent vehicular network. On the other hand,
archive an excellent overall performance than state-of-the-art the network density also changes with moving vehicles. The
works. Graph theory and algorithms have provided essential network density may be very high when traffic jam occurred,
theory support for vehicular networks in resource allocation while network density will be relatively low in suburban traffic.
and communication modeling [28], and graph applications This changing density may lead to a high communication
also help the network more informed and data-driven deci- delay. How to reduce the delay for a changing density network
sions. However, how to use graph theory and algorithms to is another challenge. At last, the distance between the vehicles
support the distinctive characteristics of intelligent vehicular is also challenging with the vehicle moves. How to update the
networks and provide high performance and real-time decision computation results, such as the shortest path between two
making policy remains challenging and represents a promising vehicle is also a challenge for existed vehicular networks.
research direction.
In this paper, we discuss some major challenges in sup- B. High Reliability and Scalability Requirements
porting intelligent vehicular networks with high performance, With the development of vehicle technologies and big
such as real-time decision making and network topologies data processing technologies, there are more and more smart
updating in dynamic changing environments, path planning, transportation and aided driving applications have been used
and QoS in high dimensional communication links. To address in our daily life. While all transportation and driving-related
these challenges, we introduce some high-performance graph applications are safety-sensitive, hence, how to design a highly
processing frameworks into intelligent vehicular network data reliable intelligent vehicular communication networks is nec-
processing. Furthermore, we proposed a fundamental shift essary. However, due to the complex network topology with

Authorized licensed use limited to: PSG COLLEGE OF TECHNOLOGY. Downloaded on July 06,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
4728 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 23, NO. 5, MAY 2022

poor stability and large network scale, how to design a highly A. Path Planning
reliable vehicular network is a great challenge. Traversal algorithms can be leveraged to find the paths in
As mentioned earlier, there are more and more vehicles dynamical vehicular networks, including the path planning and
and some other devices added into the intelligent vehicular vehicle trajectory prediction. The path planning can be further
network, which make the network proliferates. How to keep used towards navigation planning and data pre-processing for
the high scalability to keep the high quality of services (QoS) system performance improvement. For example, the single
of the network is another challenge. source shortest path (SSSP) algorithm is widely used for
path planning, and betweenness centrality (BC) is used to
C. The Potential of Graph Processing solve the traffic assignment problem [41]. Puzis et al. [41],
shows the applicability of betweenness centrality and certain
Graph processing has been widely used in kinds of appli- augmented types of it for obtaining traffic flows through links
cations to deal with the complex relations, such as molecule of a transportation network. In Rami’s work, the authors
structures model in chemistry, cognitive processes in computa- achieved a strong positive correlation with the traffic flows in
tional neuroscience, atoms structure analysis in physics, code transportation networks by using the betweenness centrality
structures analysis in computer science, and so on. To facilitate measurement and then proposed a betweenness-driven traffic
the development of large scale graph processing, there are assignment model to optimize the position of the transport
kinds of generic graph programming, and both academic and network. The proposed method can archive an excellent per-
industry searchers have implemented computing frameworks. formance due to betweenness centrality simultaneously con-
Both single machine graph processing frameworks, such as siders all shortest paths between an origin and a destination.
GraphChi [29], X-Stream [30], and GridGraph [31], and Akabane et al. [42], proposed a distributed approach to com-
distributed graph processing systems, such as Pregel [32] and pute egocentric betweenness scores over VANETs by only
PowerGraph [33] are included. use the local knowledge of the network topology, which can
Recently, the technical advance of kinds of the new hard- release the computation capacity of vehicles.
ware and accelerator, such as the General-Purpose Graphics
Processing Units (GPGPUs) [34] and FPGA, has attracted
many researchers from both academic and industry to investi- B. Vehicle Scheduling
gate how to use these kinds of new hardware and accelerators Vehicle scheduling is one of the most typical applications
to accelerate the computational and memory-intensive appli- of intelligent vehicles and smart cities, which is widely used
cations, including graph processing [35], [36]. Benefit for the in intersection management and intelligent drive control. The
massive parallelism and high memory bandwidth, the GPGPU most classic application of intersection management is to
has been widely used for graph processing frameworks. With decide the passing order of the vehicles passing through the
the efforts from both academics and industry, a set of general intersection without traditional traffic signals. The naïve imple-
graph processing frameworks have been developed, such as mentation of vehicle scheduling is the greedy method, which
Totem [37], Medusa [38], CuSha [39], and GunRock [40]. called as First-Come-First-Serve (FCFS) approach. In this
In particular, graph processing, one of the domain-specific method, the manager schedules the vehicles according to
big data processing systems, can interact with a dynamic the arrival time. The earlier come ones passing through the
environment and develop satisfactory policies to meet diverse intersection eariler. The FCFS is easy to implement, however,
data format and QoS requirements to of vehicular network in this method ignores some essential information, such as the
a dynamic and varying wireless environment. For example, interactions between vehicles and conflict zones, and thus
a vehicle adds or leaves the network will make the network leads to extra delay in many cases. On the other hand, the
structure changes. FCFS method do not take the priority into consideration,
which can delay some important vehicles and also will lead to
some serious problem. Based on improved visibility graphs,
III. G RAPH A LGORITHMS FOR IVN A PPLICATIONS
Wei [43] proposed a local path planning algorithm for an
There are kinds of graph processing algorithms that can be intelligent vehicle on a structured road. This paper introduced
used to accelerate IVN applications. For example, Breadth- the graph processing model for vehicle scheduling, which can
First Search (BFS) is widely used in kinds of graph searching improve the scheduling efficiency.
applications, Betweenness Centrality (BC) is used to find In intersection management, our objective is to minimize
important paths and vertices in IVN, some other algorithms the total time needed for all vehicles to go through the
such as Connected Component (CC) is used to find the intersection, equivalent to the leaving time of the last vehicle.
component of vehicles which with frequent communications, To remove cycles while considering the edge costs, finding a
Single Source Shortest Path (SSSP) is used find the suitable minimum spanning tree (MST) of the graph can be a poten-
paths in root/road planning, and PageRank (PR) is used to tial solution, and one approach is Kruskal’s algorithm [44].
find the important vehicle in the components of intelligent Kruskal’s algorithm repeatedly chooses a minimum-cost edge,
vehicular network applications. This section try to discuss which does not form any cycle with those already-chosen
some existed efforts to accelerate specific graph algorithms edges. Kruskal also proposed the backward version of the
on GPU. We will introduce some typical algorithms in the original one, and it repeatedly removes a maximum-cost edge
following sections. whose removal does not disconnect the graph. Inspired by this

Authorized licensed use limited to: PSG COLLEGE OF TECHNOLOGY. Downloaded on July 06,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
ZHENG AND BASHIR: GRAPH-ENABLED INTELLIGENT VEHICULAR NETWORK DATA PROCESSING 4729

method, we do intend to remove the edge, which results in the Addressing (UVA) technology is also used to allocate the
most significant delay to the objective. This can remove cycles memory space for GraphReduce, combined with the DMA
and benefit the objective minimization at the same time. Based memory loading/storing technology, GraphReduce can make
on the graph model, Lin develops a centralized cycle removal the memory access in a sequential manner, and the commu-
algorithm for the graph-based model to schedule vehicles to nication overhead can be overlapped with GPU computations
go through the intersection safely (without collisions) and through pre-fetching.
efficiently without deadlocks [45]. The algorithm is sufficiently In GTS [50], the graph data have been distinguished accord-
efficient to consider more conflict zones and more vehicles in ing to the attribute of the data, the attribute data was copied
real-time. into GPU on broad memory with a long lifetime, while the
topology data are copied to the GPU device from the host
IV. G RAPH P ROCESSING F RAMEWORKS ON GPU memory by using CUDA streaming. In this processing method,
the GPU device can quickly visit the whole graph, while
Many graph processing frameworks are proposed in recent
the extraordinary computation data can be loaded when the
years, including some single machine framework [29] and
computation moves on. GTS also introduced the SSDs to
distributed frameworks [33], and also some frameworks
accelerate the data transfer operation.
on new hardware devices, such as GPUs [40] and FPGA
[46], [47]. In this paper, we consider the character of the
IVN data, and try to conclude some optimization strategies for C. Workload Mapping
graph processing framework for IVN data processing. In this The workload imbalance problem is caused by the irregular
section, we discuss the programming models and system vertex degree. How to mapping the uneven workload onto
implementation and optimization technologies. GPU is another challenge. The mostly used method is evenly
assign the thread to the edges. In this method, every thread
A. Graph Programming Models just processes an edge, the computation task of every thread
is the same. However, there are much more threads are
The GAS (Gather-Apply-Scatter) and BSP (Bulk Syn-
needed than assign threads to vertices, which will slow down
chronous Parallel) programming models are mostly used,
the overall performance. In order to solve this problem,
BSP is firstly used in Pregel [32] and GAS is proposed in
Dynamic scheduling strategy is proposed in MapGraph [48].
PowerGraph [33].
In MapGraph the threads are assigned according to the vertex’s
BSP programming model is widely used in many graph
degree, a warp is assigned to the vertices with lower degree
processing frameworks on GPU, including TOTEM, Medusa
while the CTA is assigned the vertices with higher degree.
and GunRock. In BSP programming model, the program
In this thread assignment model, it is easy to distinguish
executed as a so-called super-steps. The threads run asynchro-
different workloads, and then it can dynamically schedule the
nously in parallel within a super-step, while all the threads
workload.
need to be synchronized at the end of the super-step. Hence,
Gunrock [40] implemented a similar thread assignment
the threads only can communicate with each other at the end
fashion with MapGraph [48], but Gunrock classified the tasks
of the super-steps, which is called a barrier. This execution
into fine- and coarse-grained workload, then assign the thread
fashion is easy to implement, but it is easy to lead the straggler
warp/block to the different workloads. A virtual warp include
problem since the computation task on the vertices are varies.
multiple warps is also implemented in Gunrock, which is
There are also many graph processing frameworks on GPU
assigned to the vertex with tremendous degree.
are adopt the GAS model, such as MapGraph [48] and
CuSha [39]. In the GAS programming model, the program
on each vertex is divided into three phases, which are named D. Miscellaneous
Gather, Scatter, and Apply. In Gather phase, the vertices col- There are some other research challenges in graph-
lect information from their neighbors, and the updates from the enabled IVN data processing, for example, how to provide a
neighbors will be added in the Apply phase, then the Scatter programmer-friendly interface and how to improve the degree
phase broadcast the updates to the neighbors to promotes the of parallelism.
program into the next iteration. In this fashion, the GAS can Medusa [38] and MapGraph [48] both provides a set of
be implemented in an asynchronous fashion, which can avoid programmer-friendly APIs, which is easy for the programmers
the synchronization overhead and the straggler problem. to implement the user-defined functions. In addition, Medusa
also provides a set of configuration parameters to enhance the
B. Memory Access Pattern flexibility of the framework. Unlike the previous implementa-
tion, GunRock provides a data-centric abstraction to enhance
The random memory access manner of graphs will limit
the flexibility of the library.
the performance of the GPU, how to unleash the computation
ability of GPU is a great challenge. In this section, we discuss
some existed technologies on memory access optimization. V. G RAPH -E NABLED IVN F RAMEWORK
In GraphReduce [49], the vertices are sorted according to Graph processing represents a useful tool for a variety of
the source vertices ID to make sure all the edge from the applications. In this paper, we proposed a graph-enabled IVN
same vertex can be visited at the same time. A Unified Virtual data processing framework, Hooker. In Hooker, we designed

Authorized licensed use limited to: PSG COLLEGE OF TECHNOLOGY. Downloaded on July 06,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
4730 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 23, NO. 5, MAY 2022

Fig. 2. The JCS programming model and how it matches with multiple
hardware.
Fig. 3. Data partitioning and placement strategies for data-aware and
structure-aware.

a new programming model to fit multi-architectures, a new

communication paradigm to maintain the data consistency, task size, including the for-all parallelism, reduction
integrality and security. parallelism, and scan parallelism, it also supports the
irregular parallelism, which can achieves excellent load
A. The JCS Programming Model balancing through reasonable task stealing scheduling.
• Heuristic parallel writes optimization mechanism.
Vertex-centric programming model is easy to express most There are many duplicate vertices in the worklist since
of the graph algorithms, and it can provide high scalability there are more than one vertices connected with the
by partition the graphs. However, it is easy to lead random same vertex. Atomic operations and locks are intro-
memory access and load imbalance problem, due to the duced to ensure data consistency, but both atomic oper-
skewed degree distribution of real-word graphs. Edge-centric ations and lock operation will lead lost of conflicts
programming model can provide a continuous memory access with will slow down the overall performance. However,
fashion, but it will introduce lots of redundant computation recent researches show that some updating operations
since there are many more edges than the vertices in real- are idempotent (i.e. the updating order does not affect
world graphs. The Gather-Apply-Scatter (GAS) is a fine- data consistency). Hence, there is no need to design a
grained vertex-centric programming model, which proposed specific algorithm by using atomic operation or lock to
in PowerGraph [33]. Previous research shows that the num- remove the duplicate vertices from the worklist, design a
ber of active vertices in each iteration is far less than the lightweight runtime heuristic policy is enough to remove
total number of vertices in a graph [33]. Hence, this paper the redundant computation, which will be more efficient.
proposes a queue-based vertex-centric Join-Compute-Scatter
(JCS) programming model, which is shown in figure 2. In JCS
programming model, the operation on the vertex is divided B. Feature-Aware Data Partitioning and Placement Strategy
into join, compute and scatter three steps. The join operation In order to meet the architecture’s feature to unleash the
adds the active vertices into the worklist, and the compute device performance, this paper propose a three-dimensional
operation updates the vertex’s value according to the user- graph partition scheme. The proposed graph partition scheme
defined function, while the scatter operation scatter the vertex’s will load the graph blocks into the device on broad memory
value to its neighbours, which similar like the scatter operation to make sure the locality of data access and hence to reduce
in GAS model. In JCS model, each iteration cares about the the I/O overhead by considering the communication overhead.
vertices which need to be updated. This execution fashion Figure 3 show the essential operation of the proposed graph
can provide a unified and concise implementation for different partition scheme.
algorithms, and it can provide high scalability for the vertex- The existed graph processing system applied the one-
centric method. dimensional or two-dimensional partition strategy, which is the
In order to make the JCS model matches with different kinds vertex-centric and edge-centric partition method, respectively.
of hardware architecture, this paper provides a mixed granu- The vertex-centric partition strategy will lead to a load imbal-
larity task mapping mechanism and a heuristic parallel write ance problem, since the vertex degree distribution of most real-
optimization mechanism. We introduce the two optimization world graphs is power-law. While the edge-centric partition
mechanisms as follow: strategy will lead to a large amount of communication between
• Mixed-granularity task mapping mechanism. Here, the master node and the replicas. Recent research proposed a
we take GPU as an example to introduce the mixed- 3D partition strategy, which partition the vertex attribution as
granularity task mapping mechanism. We assign the the third-dimensional partition object, this partition can archive
vertices to thread, warp, CTA, and kernel according to an excellent system performance in machine learning applica-
the number of the neighbour of the active vertex, and tions, but do not suitable for full broad applications [31]. The
the virtual-warp is used to solve the load imbalance traversal tree-based partition method can maintain good local-
problem. While, on KNL, the proposed mechanism not ity, but the partition operation executed after traverse the whole
only supports regular round-level parallelism for different graph. The overhead of this partitioning method is enormous.

Authorized licensed use limited to: PSG COLLEGE OF TECHNOLOGY. Downloaded on July 06,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
ZHENG AND BASHIR: GRAPH-ENABLED INTELLIGENT VEHICULAR NETWORK DATA PROCESSING 4731

In order to solve the problems of the existed partition methods,

this paper proposes a hybrid partition method. The proposed
method first partition the vertices with similar degrees together
by using a two-dimensional partition method and then partition
the sub-graphs again by using a one-dimensional partition
method. The hybrid partition strategy will take the third-
dimensional partition as the first round partition for some
specific graph algorithms, and then partition the results again
by using a two-dimensional partition method. This method can
archive load balance and very low communication computation
ratio for different algorithms and architectures.
The data placement is closely related to the representation of Fig. 4. Layer centric communication model.
the graph, and it has a deep effect on system performance. The
upper-level framework requires the graph representation with
high memory bandwidth utilization and locality of memory and what kind of framework we can use to process the different
access. While, the lower-level storage requires high space types of intelligent vehicular network data.
utilization, avoid the space-wasting for sparsity graph. The
storage level also requires the graph can be loaded into the
A. Experimental Datasets and Algorithms
memory during the I/O operation in graph processing. In order
to meet both the upper-level and lower-level’s requirements, In this section, we introduce some typical datasets and
this paper provides a hybrid CSR/CSC graph representation, algorithms which chosen to demonstrate the performance of
and the edge-list representation is also provided according to the graph-enabled IVN data processing frameworks.
the characteristics of the graph. The mixed CSR/CSC graph All the datasets of the experiments conducted in this paper
representation is benefit for the Scatter/Gather operation. For are follow the classic graph formalism [51]. We use V and
example, some implementation of the BFS algorithm will E to present the vertices and edges of the graph, respectively.
change the traversal direction from bottom-up to top-down G = (V, E) present the graph. The edge presented as e, where
(or top-down to bottom-up). This hybrid representation will e = (u, v) and e =< u, v > are the undirected and directed
improve memory access efficiency by changing CSR to CSC edges. In this paper, both directed and undirected graphs used
(or change CSC to CSR). Community is another essential in our experiments.
characteristic of the real world graphs, i.e. the vertices in a In order to include both power-law and large diameter
community are connected but few vertices connect with the graphs, we select six graphs from different real-world appli-
vertices outside the community. In this kind of graph, the cations, such as e-business, social network and some other
community can be processed by some SIMD devices, such source networks. All these graphs are with different structures
as GPU, by using the edge-list representation will archive an and a varying number of vertices and edges. The graphs are
excellent performance. Hence, edge-list representation can be shown as table I. All these six graphs can be downloaded
an optional method for users. from the Stanford Network Analysis Project (SNAP) [52].
As the power-law graph follow a distribution, shown as
formula 1 [53], [54], we list the exponent and the fitness in
C. Comunnications
table I to compare the power-law attribution. In this paper, the
In order to processing large scale IVN graphs, we proposed graphs stored in a plain text file with edgelist format, which
a layer centric communication model for Hooker, which is is easy for us to locate the edges.
shown in figure 4. In this communication model, the threads
in each layer passing messages to each other asynchronous, P(x) ∝ x −α (1)
while the messages need to be synchronized between layers. In order to compare the performance of different algorithms
This communication model can accelerate the task in each on all these frameworks, we implemented four most frequently
layer and keep the data consistency, integrality and security used graph algorithms, include two traversal algorithms (BFS
between layers. and SSSP) and two iterative algorithms (PageRank and Con-
nected Component).
VI. E XPERIMENTS All the experiments are conducted on an NVIDIA GeForce
In this paper, we conducted a set of experiments to try GTX 1060 GPU, which is a Maxwell architecture device
to find some insights for graph enabled IVN data process- with 1280 CUDA cores and 6GB onboard memory. The
ing framework. On the one hand, we compare the perfor- programs compiled by using CUDA 8.0 with “-arch=sm 35”
mance of graph processing frameworks with different types of flag on Ubuntu 16.04. We also compared the graph frameworks
graphs, which widely used in an intelligent vehicular network. on GPU with some frameworks which run on CPU, such as
On the other hand, we compare the performance of different GraphChi [29], X-Stream [30], and GridGraph [31]. All the
graph processing frameworks, which implemented on different CPU enabled graph frameworks are run on a machine with
devices. Hence, we can conclude that when we need to use 8GB memory and 4 Intel Core(TM) i5-5200 CPUs at 2.2GHz.
some accelerator device to accelerate the algorithm on GPU All the source codes are provided by the authors.

Authorized licensed use limited to: PSG COLLEGE OF TECHNOLOGY. Downloaded on July 06,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
4732 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 23, NO. 5, MAY 2022

TABLE I
D ATASETS U SED IN THE E XPERIMENTS

B. Graph-Processing Frameworks
Both graph processing frameworks on GPU and CPU,
which are widely used in research and industrial communities,
are compared. The famewroks on GPU are TOTEM [37],
CuSha [39], Medusa [38], Gunrock [40], and MapGraph [48].
TOTEM is the first GPU and CPU hybrid system [37],
which partition the graphs into two parts, one partition is
processed on GPU and the other one is processed on CPU.
The vertex-centric processing pattern is used in TOTEM, and
there are three different partition strategies for TOTEM, named
HIGH-degree, LOW-degree and RANDOM-degree, according
to the vertices’ degree. The HIGH-degree partition strategy
assigns the high degree vertices to the CPU and assigns the
low degree vertices to the GPU. LOW-degree partition strategy
is opposite with the HIGH-degree strategy, and the RANDOM-
degree strategy the vertices to CPU and GPU device in a
random manner. Medusa and CuSha adopt the edge-centric
processing model, G-Shards and Concatenated Windows (CW)
technology like GraphChi are implemented in CuSha, while
Medusa provides a set of simplified programming interface.
MapGraph implemented the GAS strategy, which proposed
in PowerGraph, on GPU. Gunrock is the most recent high-
performance library on GPUs. All the parameters in this paper
are the same with the authors’ original work list in their
publication, which can achieve the best system performance.

C. Experiment Results
1) Perspective of Dataset: We first try to conclude that
types of datasets and algorithms which can executed on GPU. Fig. 5. System performance for different graphs on CPU and GPU.
In this experiment, the BFS is implemented as a textbook
version, while PR and SSSP are implemented according to
the best-reported performance in [37], and CC is implemented with the host through PCI-E bus, while the bandwidth of
according to Wu and Du [55] version. Figure 5 shows the PCI-e is limited. Hence, the compunction cost is the main
experimental performance of different algorithms executed bottleneck of PageRank on GPU, in particular for the large
different datasets for both GPU and CPU. scale graphs which with the graph size larger than the GPU
Figure 5 shows that the algorithms achieve better perfor- on broad memory. The experiment also shows that the power-
mance on GPU than on CPU except for PageRank. As we law datasets, such as Amazon and Twitter can achieve a
discussed before, GPU has higher computing power than CPU good speed up when the dataset can be loaded into the GPU
when the memory access is regular. However, a graph is of on broad memory, while the large diameter graphs such as
the irregular data structure, and the GPU memory accessed RoadNet can not achieve such an excellent performance.
irregularly in graph algorithms. Figure 5 also indicts that GPU Comparing Figure 5(a), Figure 5(c) and Figure 5(f), we can
is hard to process the large scale graph, particularly when the conclude that the algorithms performance on Amazon and
graph size reaches the memory size. The experiment shows Twitter outperforms on RoadNet. This result indicates that
the performance of PageRank on GPU is even worse than GPU is more suitable to process power-law graphs than CPU.
CPU, this is because PageRank needs to send the updates 2) Perspective of Algorithms: In order to investigate the
of the vertices to their neighbours in every iteration, there performance of different algorithms, we run the selected
are many communication operations. As the GPU connected algorithms on different GPU graph processing systems.

Authorized licensed use limited to: PSG COLLEGE OF TECHNOLOGY. Downloaded on July 06,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
ZHENG AND BASHIR: GRAPH-ENABLED INTELLIGENT VEHICULAR NETWORK DATA PROCESSING 4733

TABLE II
E XECUTION T IME OF BFS ( MS )

TABLE III
E XECUTION T IME OF PAGE R ANK ( MS )

TABLE IV
E XECUTION T IME OF CC ( MS )

Fig. 6. The performance of different algorithms on CPU graph processing

system.

TABLE V
E XECUTION T IME OF SSSP ( MS ) for random access to vertices. We can see that CPU-based
graph processing systems achieve higher performance than the
GPU-based systems when processing the PageRank algorithm.
This is because PageRank uses a considerable number of
random accesses, which will slow down the GPU performance.
From Table II–V we can conclude that not all the graph
algorithms are suitable for the GPU graph processing system.
For each combination of algorithm and graph processing BFS-like algorithms, which access the memory regularly by
framework, we run the algorithm 5 times and obtain the using some data layout techniques, are suitable to be run by
average performance. The experimental results are shown in the GPU-based graph processing system, so are the power-law
Tables II–V. graphs.
We can conclude from the experimental results that Gunrock
outperforms all other GPU-based graph processing systems VII. C ONCLUSION AND F UTURE O PPORTUNITIES
with BFS. This is because Gunrock makes use of the frontier The IVN is the basis of smart cities, how to processing the
to store its graph data, creating a ‘queue’ for the data structure. dynamic IVN data is a significant challenge. Nowadays, most
This works well with BFS-like algorithms. TOTEM follows of the big data processing systems are under a batch processing
Gunrock as the second most competitive system. model, which is hard to process the dynamic data. The graph
Table II shows a comparison of the performance of all processing frameworks, with a set of data processing paradise
six GPU based graph processing systems on the BFS algo- for the highly connected data, may provide a chance for
rithm. MapGraph has a clear performance advantage on IVN data processing. This paper first discussed the research
Amazon, however, Medusa outperforms other systems on challenges of IVN data processing and then proposed a new
DBLP. RoadNet-CA is the dataset with the biggest diameter, paradise for IVN data processing.
which makes the processing more difficult for BFS. However, In the future, we plan to focus on some other aspects of
MapGraph and Gunrock achieve good performance with this IVN data processing, such as the IVN processing framework
dataset. This is because they use the queue-like data structure. on some other new devices like Directed Self Assembly (DSA)
The damping factor and precision of PageRank in our and Field Programmable Gate Array (FPGA), the real-time
experiments set as 0.85 and 0.005, respectively. Table III dynamic data processing for IVN or some other new devices.
shows that CuSha performs better than any other system
with PageRank, which mainly benefits from the strategy of R EFERENCES
G-shards partition. [1] L. Liang, H. Peng, G. Y. Li, and X. Shen, “Vehicular communications:
We also looked at three CPU-based graph processing frame- A physical layer perspective,” IEEE Trans. Veh. Technol., vol. 66, no. 12,
works, GraphChi, X-Stream, and GridGraph. GraphChi chops pp. 10647–10659, Dec. 2017.
[2] G. Araniti, C. Campolo, M. Condoluci, A. Iera, and A. Molinaro, “LTE
large graphs into small parts. It designs a Parallel Sliding for vehicular networking: A survey,” IEEE Commun. Mag., vol. 51,
Windows (PSW) method in order to gain a better efficiency no. 5, pp. 148–157, May 2013.

Authorized licensed use limited to: PSG COLLEGE OF TECHNOLOGY. Downloaded on July 06,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
4734 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 23, NO. 5, MAY 2022

[3] W. Wu, Z. Yang, and K. Li, “Internet of vehicles and applica- [24] C. L. Azevedo, J. L. Cardoso, and M. Ben-Akiva, “Vehicle tracking
tions,” in Internet Things, R. Buyya and A. V. Dastjerdi, Eds. using the k-shortest paths algorithm and dual graphs,” Transp. Res.
San Mateo, CA, USA: Morgan Kaufmann, 2016, pp. 299–317. Proc., vol. 1, no. 1, pp. 3–11, 2014. [Online]. Available: [Link]
[Online]. Available: [Link] [Link]/science/article/pii/S2352146514000039
B9780128053959000162 [25] S. Dietzel, J. Petit, G. Heijenk, and F. Kargl, “Graph-based metrics
[4] E. Ahmed and H. Gharavi, “Cooperative vehicular networking: A for insider attack detection in VANET multihop data dissemination
survey,” IEEE Trans. Intell. Transp. Syst., vol. 19, no. 3, pp. 996–1014, protocols,” IEEE Trans. Veh. Technol., vol. 62, no. 4, pp. 1505–1518,
Mar. 2018. May 2013.
[5] B. Menezes, C. Canales, T. Zimmerman, and M. Toussaint. Gart- [26] M. H. Eiza and Q. Ni, “An evolving graph-based reliable routing scheme
ner 2019 Magic Quadrant for the Wired and Wireless LAN for VANETs,” IEEE Trans. Veh. Technol., vol. 62, no. 4, pp. 1493–1504,
Access Infrastructure. Accessed: Dec. 20, 2019. [Online]. Available: May 2013.
[Link] [27] K. Zheng, F. Liu, Q. Zheng, W. Xiang, and W. Wang, “A graph-based
[6] X. Cheng, L. Yang, and X. Shen, “D2D for intelligent transportation cooperative scheduling scheme for vehicular networks,” IEEE Trans.
systems: A feasibility study,” IEEE Trans. Intell. Transp. Syst., vol. 16, Veh. Technol., vol. 62, no. 4, pp. 1450–1458, May 2013.
no. 4, pp. 1784–1793, Aug. 2015. [28] H. Kröner, “Radio resource allocation for data services in UMTS
[7] L. Zhu, F. R. Yu, Y. Wang, B. Ning, and T. Tang, “Big data analytics in networks,” AEU Int. J. Electron. Commun., vol. 55, no. 1, pp. 55–62,
intelligent transportation systems: A survey,” IEEE Trans. Intell. Transp. Jan. 2001. [Online]. Available: [Link]
Syst., vol. 20, no. 1, pp. 383–398, Jan. 2019. article/pii/S1434841104702472
[8] J. Seo and T. Walter, “Future dual-frequency GPS navigation system [29] A. Kyrola, G. Blelloch, and C. Guestrin, “Graphchi: Large-scale
for intelligent air transportation under strong ionospheric scintillation,” graph computation on just a PC,” in Proc. 10th USENIX Symp.
IEEE Trans. Intell. Transp. Syst., vol. 15, no. 5, pp. 2224–2236, Operating Syst. Design Implement. (OSDI), Hollywood, CA, USA:
Oct. 2014. USENIX, 2012, pp. 31–46. [Online]. Available: [Link]
[9] J. B. Kenney, “Dedicated short-range communications (DSRC) standards conference/osdi12/technical-sessions/presentation/kyrola
in the United States,” Proc. IEEE, vol. 99, no. 7, pp. 1162–1182, [30] A. Roy, I. Mihailovic, and W. Zwaenepoel, “X-stream: Edge-
Jul. 2011. centric graph processing using streaming partitions,” in Proc. 24th
[10] M. Jutila, J. Scholliers, M. Valta, and K. Kujanpää, “ITS-G5 perfor- ACM Symp. Operating Syst. Princ., Nov. 2013, pp. 472–488, doi:
mance improvement and evaluation for vulnerable road user safety 10.1145/2517349.2522740.
services,” IET Intell. Transp. Syst., vol. 11, no. 3, pp. 126–133, 2017. [31] X. Zhu, W. Han, and W. Chen, “Gridgraph: Large-scale graph
[11] M. Wegner, T. Schwarz, and L. Wolf, “Connectivity maps for V2I processing on a single machine using 2-level hierarchical partitioning,”
communication via ETSI ITS-G5,” in Proc. IEEE Veh. Netw. Conf. in Proc. USENIX Annu. Tech. Conf. (USENIX ATC), Santa Clara,
(VNC), Dec. 2018, pp. 1–8. CA, USA: USENIX Association, Jul. 2015, pp. 375–386. [Online].
Available: [Link]
[12] IEEE Standard for Information Technology—Local and Metropolitan
presentation/zhu
Area Networks—Specific Requirements—Part 11: Wireless LAN Medium
[32] G. Malewicz et al., “Pregel: A system for large-scale graph process-
Access Control (MAC) and Physical Layer (PHY) Specifications Amend-
ing,” in Proc. ACM SIGMOD Int. Conf. Manage. data, Jun. 2010,
ment 6: Wireless Access in Vehicular Environments, Standard 802.11p-
pp. 135–146, doi: 10.1145/1807167.1807184.
2010 (Amendment to IEEE Std 802.11-2007 as amended by IEEE Std
802.11k-2008, IEEE Std 802.11r-2008, IEEE Std 802.11y-2008, IEEE [33] J. E. Gonzalez, Y. Low, H. Gu, D. Bickson, and C. Guestrin, “Pow-
Std 802.11n-2009, and IEEE Std 802.11w-2009), Jul. 2010, pp. 1–51. ergraph: Distributed graph-parallel computation on natural graphs,”
in Proc. 10th USENIX Symp. Operating Syst. Design Implement.
[13] G. Araniti, C. Campolo, M. Condoluci, A. Iera, and A. Molinaro, “LTE
(OSDI). Hollywood, CA, USA: USENIX, 2012, pp. 17–30. [Online].
for vehicular networking: A survey,” IEEE Commun. Mag., vol. 51,
Available: [Link]
no. 5, pp. 148–157, May 2013.
presentation/gonzalez
[14] A. Vinel, “3GPP LTE versus IEEE 802.11p/WAVE: Which tech-
[34] J. D. Owens et al., “A survey of general-purpose computation on
nology is able to support cooperative vehicular safety applica-
graphics hardware,” Comput. Graph. Forum, vol. 26, no. 1, pp. 80–113,
tions?” IEEE Wireless Commun. Lett., vol. 1, no. 2, pp. 125–128,
Mar. 2007, doi: 10.1111/j.1467-8659.2007.01012.x.
Apr. 2012.
[35] D. Merrill, M. Garland, and A. Grimshaw, “Scalable GPU graph
[15] A. Sultan and M. Pope, “3rd generation partnership project; technical traversal,” in Proc. 17th ACM SIGPLAN Symp. Princ. Pract. Parallel
specification group services and system aspects; release 14 description; Program. (PPoPP), 2012, pp. 117–128, doi: 10.1145/2145816.2145832.
summary of REL-14 work items(release 14),” 3GPP, Valbonne, France,
[36] A. Ashari, N. Sedaghati, J. Eisenlohr, S. Parthasarath, and
Tech. Rep., 21.914, 2018.
P. Sadayappan, “Fast sparse matrix-vector multiplication on GPUs
[16] L. Liang, S. Xie, G. Y. Li, Z. Ding, and X. Yu, “Graph-based radio for graph applications,” in Proc. Int. Conf. High Perform. Comput.,
resource management for vehicular networks,” in Proc. IEEE Int. Conf. Netw., Storage Anal., (SC), Nov. 2014, pp. 781–792.
Commun. (ICC), May 2018, pp. 1–6. [37] A. Gharaibeh, L. B. Costa, E. Santos-Neto, and M. Ripeanu, “A yoke
[17] Z. Li, J. Liu, and Y. Chen, “An energy-efficient resource allocation of oxen and a thousand chickens for heavy lifting graph processing,” in
strategy for vehicular networks,” in Proc. ACM Turing Celebration Conf. Proc. 21st Int. Conf. Parallel Archit. Compilation Techn. (PACT), 2012,
China, May 2019, pp. 1–9, doi: 10.1145/3321408.3326657. pp. 345–354, doi: 10.1145/2370816.2370866.
[18] D. Liu, M.-C. Lee, C.-M. Pun, and H. Liu, “Analysis of wireless [38] J. Zhong and B. He, “Medusa: Simplified graph processing on GPUs,”
localization in nonline-of-sight conditions,” IEEE Trans. Veh. Technol., IEEE Trans. Parallel Distrib. Syst., vol. 25, no. 6, pp. 1543–1552,
vol. 62, no. 4, pp. 1484–1492, May 2013. Jun. 2014.
[19] Y. Huang, M. Chen, Z. Cai, X. Guan, T. Ohtsuki, and Y. Zhang, “Graph [39] F. Khorasani, K. Vora, R. Gupta, and L. N. Bhuyan, “CuSha: Vertex-
theory based capacity analysis for vehicular ad hoc networks,” in Proc. centric graph processing on GPUs,” in Proc. 23rd Int. Symp. High-
58th IEEE Global Commun. Conf. (GLOBECOM), San Diego, CA, Perform. Parallel Distrib. Comput. (HPDC), 2014, pp. 239–252, doi:
USA, Dec. 2015, pp. 1–5. 10.1145/2600212.2600227.
[20] K. Zhang, L. Lan, J. Kwok, S. Vucetic, and B. Parvin, “Scaling up [40] Y. Wang, A. Davidson, Y. Pan, Y. Wu, A. Riffel, and J. D. Owens,
graph-based semisupervised learning via prototype vector machines,” “Gunrock: A high-performance graph processing library on the GPU,”
IEEE Trans. Neural Netw. Learn. Syst., vol. 26, no. 3, pp. 444–457, in Proc. 21st ACM SIGPLAN Symp. Princ. Pract. Parallel Program.
Mar. 2015. (PPoPP), New York, NY, USA: ACM, 2016, pp. 11:1–11:12, doi:
[21] M. Nickel, K. Murphy, V. Tresp, and E. Gabrilovich, “A review of 10.1145/2851141.2851145.
relational machine learning for knowledge graphs,” Proc. IEEE, vol. 104, [41] R. Puzis, Y. Altshuler, Y. Elovici, S. Bekhor, Y. Shiftan, and A. Pentland,
no. 1, pp. 11–33, Jan. 2016. “Augmented betweenness centrality for environmentally aware traffic
[22] Z. Zhang, P. Cui, and W. Zhu, “Deep learning on graphs: A survey,” monitoring in transportation networks,” J. Intell. Transp. Syst., vol. 17,
IEEE Trans. Knowl. Data Eng., vol. 34, no. 1, pp. 249–270, Jan. 2022. no. 1, pp. 91–105, Jan. 2013, doi: 10.1080/15472450.2012.716663.
[23] C. L. Azevedo, J. L. Cardoso, and M. Ben-Akiva, “Applying graph [42] A. T. Akabane, R. Immich, R. W. Pazzi, E. R. M. Madeira, and
theory to automatic vehicle tracking by remote sensing,” in Proc. 93rd L. A. Villas, “Distributed egocentric betweenness measure as a vehi-
Annu. Meeting Transp. Res. Board (TRB), Washington, DC, USA: TRB, cle selection mechanism in vanets: A performance evaluation study,”
2014, pp. 1–12. [Online]. Available: [Link] Sensors, vol. 18, no. 8, pp. 1–27, 2018.

Authorized licensed use limited to: PSG COLLEGE OF TECHNOLOGY. Downloaded on July 06,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
ZHENG AND BASHIR: GRAPH-ENABLED INTELLIGENT VEHICULAR NETWORK DATA PROCESSING 4735

[43] H. Wei, “Path planning algorithm of intelligent vehicle based on [54] Y. Virkar and A. Clauset, “Power-law distributions in binned empir-
improved visibility graphs,” in Proc. Intell. Connected Vehicles Symp., ical data,” Ann. Appl. Statist., vol. 8, no. 1, pp. 89–119, 2014, doi:
Aug. 2018, pp. 1–8, doi: 10.4271/2018-01-1581. 10.1214/13-AOAS710.
[44] J. B. Kruskal, Jr., “On the shortest spanning subtree of a graph [55] B. Wu and Y. Du, “Cloud-based connected component algo-
and the traveling salesman problem,” Proc. Amer. Math. Soc., rithm,” in Proc. Int. Conf. Artif. Intell. Comput. Intell., Oct. 2010,
vol. 7, no. 1, pp. 48–50, Feb. 1956. [Online]. Available: [Link] pp. 122–126.
[Link]/stable/2033241
[45] Y.-T. Lin, H. Hsu, S.-C. Lin, C.-W. Lin, I. H.-R. Jiang, and C. Liu,
“Graph-based modeling, scheduling, and verification for intersection
management of intelligent vehicles,” ACM Trans. Embedded Comput.
Syst., vol. 18, no. 5s, pp. 1–21, Oct. 2019, doi: 10.1145/3358221.
[46] G. Dai, Y. Chi, Y. Wang, and H. Yang, “FPGP: Graph process-
ing framework on FPGA a case study of breadth-first search,” in Zhigao Zheng (Member, IEEE) is with the School of Computer Science and
Proc. ACM/SIGDA Int. Symp. Field-Program. Gate Arrays, Feb. 2016, Technology, Huazhong University of Science and Technology. His research
pp. 105–110, doi: 10.1145/2847263.2847339. interests include parallel and distributed systems and graph processing
[47] G. Dai, T. Huang, Y. Chi, N. Xu, Y. Wang, and H. Yang, “ForeGraph: algorithms.
Exploring large-scale graph processing on multi-FPGA architecture,”
in Proc. ACM/SIGDA Int. Symp. Field-Programmable Gate Arrays,
Feb. 2017, pp. 217–226, doi: 10.1145/3020078.3021739.
[48] Z. Fu, M. Personick, and B. Thompson, “MapGraph: A high level API
for fast development of high performance graph analytics on GPUs,”
in Proc. Workshop GRAph Data Manage. Experiences Syst., Jun. 2014,
pp. 1–6, doi: 10.1145/2621934.2621936. Ali Kashif Bashir (Senior Member, IEEE) is currently a Reader with
[49] D. Sengupta, S. L. Song, K. Agarwal, and K. Schwan, “GraphReduce: the Department of Computing and Mathematics, Manchester Metropolitan
Processing large-scale graphs on accelerator-based systems,” in Proc. University, U.K. He is also affiliated with the University of Electronic Science
Int. Conf. High Perform. Comput., Netw., Storage Anal., Nov. 2015, and Technology of China (UESTC), China; the National University of Science
pp. 1–12. and Technology, Islamabad (NUST), Pakistan; and the University of Guelph,
[50] M.-S. Kim, K. An, H. Park, H. Seo, and J. Kim, “GTS: A fast Canada, in honorary roles. He is associated with a few research and industrial
and scalable graph processing method based on streaming topology to projects as a PI and Co-I, accumulatively of around seven million pounds. He
GPUs,” in Proc. Int. Conf. Manage. Data, Jun. 2016, pp. 447–461, doi: has published over 200 research articles, delivered more than 30 invited talks
10.1145/2882903.2915204. across the globe, organized more than 40 guest editorials, and chaired around
[51] D. B. West, Introduction to Graph Theory (Math Classics). 35 conferences and workshops. He is also serving as a Reviewer for EPSRC,
Upper Saddle River, NJ, USA: Prentice-Hall, 2001. U.K.; Commonwealth, U.K.; the National Science and Engineering Research
[52] J. Leskovec. (2009). Stanford Network Analysis Project. [Online]. Avail- Council (NSERC), Canada; MITACS, Canada; Irish Research Council (IRC),
able: [Link] Ireland; and Qatar National Research Fund (QNRF), Qatar. Since 2016, he
[53] A. Clauset, C. R. Shalizi, and M. E. J. Newman, “Power-law distribu- has been serving IEEE Technology, Policy and Ethics Newsletter as an Editor-
tions in empirical data,” SIAM Rev., vol. 51, no. 4, pp. 661–703, 2009, in-Cheif. He is also an Associate Editor of several IEEE, Springer, IET, and
doi: 10.1137/070710111. MDPI journals.

Authorized licensed use limited to: PSG COLLEGE OF TECHNOLOGY. Downloaded on July 06,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.