Efficient Placement of Edge Computing Devices for Vehicular Applications in Smart Cities
1. Efficient Placement of Edge Computing Devices for
Vehicular Applications in Smart Cities
Gopika Premsankar, Bissan Ghaddar, Mario Di Francesco, Rudi
Verago
2. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
2/28
Connected vehicles in smart cities
Autonomous cars
Collision avoidance
Dynamic real-time
routing of vehicles
Real-time computer
vision applications
3. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
3/28
Connected vehicles: challenges
Mobile vehicles need continuous connectivity
Low latency communication
Increasing amount of data to be processed
4. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
4/28
Edge computing in vehicular networks
Smart road side units (RSUs) equipped with servers
Co-located
Server
Cloud
Roadside
Unit
Edge Computing Device
Urban Area
V2I V2I
6. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
6/28
Objective
How to deploy a network of edge computing devices in
a city for connected cars?
Challenges:
Mobile vehicles need continuous connectivity
Dense and built-up environment
Limited computing resources at edge
7. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
7/28
Urban area deployments
8. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
8/28
Contribution
Propose a mixed integer linear programming formulation
for efficient deployment of edge computing devices
Features:
Specify target network and computational demand coverage
Accurately characterizes the effect of built-up environment on
communications
Uses open public data
9. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
9/28
Agenda
State of the art
Mixed Integer Linear Programming formulation
Evaluation
Future work
11. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
11/28
Optimal placement of RSUs
So far in the literature:
Placement strategies to ensure network coverage [Liang et
al., VTC ’12, Aslam et al., ISCC ’12, Balouchzahi et al., 2015]
Aim to reduce cost and number of RSUs deployed
Our approach:
Placement of RSUs in urban areas taking into account both
network coverage and computation requirements
Y. Liang, H. Liu, D. Rajan, “Optimal placement and configuration of roadside units in vehicular networks”, IEEE
Vehicular Technology Conference, 2012
B. Aslam, F. Amjad, C. C. Zou, “Optimal roadside units placement in urban areas for vehicular networks”, 2012
ISCC, July 2012, pp. 423-429
N. M. Balouchzahi, M. Fathy, A. Akbari, “Optimal road side units placement model based on binary integer
programming for efficient traffic information advertisement and discovery in vehicular environment”, IET Intelligent
Transport Systems, vol. 9, no. 9, pp. 851-861, 2015
12. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
12/28
System model
Candidate RSU location
Area modelled as a set of cells C
Each cell has a candidate location for installing an RSU
RSUs can be deployed with different power levels P
Wireless communication over IEEE 802.11p
13. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
13/28
Optimization problem: RSU-OPT
A mixed integer linear programming model to
minimize cost of deploying RSUs while ensuring:
target network coverage γ
target computational demand α
Decision variables:
yi Place RSU in cell i
xi,k Transmit power level k to assign to RSU in cell i
14. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
14/28
RSU-OPT: objective
minimize
|C|
i=1
cyi +
|C|
i=1
|C|
j=1
ai,j zi,j +
|P|
k=1
|C|
i=1
bk xi,k
Fixed cost for each RSU
Cost based on distance between RSU and covered cell
Cost based on transmit power level
15. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
15/28
RSU-OPT: constraints for network coverage
Pre-computed
adjacency matrix
Meet network coverage
requirement for gamma
percent of roads
16. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
16/28
RSU-OPT: constraints for computational demand
Meet computational demand
for alpha percentage of area
Do not exceed available
CPU cycles at each RSU
18. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
18/28
Evaluation: simulation settings
Area: city center of Dublin (3.2 by 3.1 square kilometers)
Under moderate and high traffic conditions
Network simulations
MILP solved using CPLEX
Key simulation parameters:
Parameter Value
Edge server 1.2 GHz
Computational availability 600 megacycles
Computational requirement 18,000 CPU cycles/bit
Application message frequency 1 Hz
19. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
19/28
Evaluation: comparison with other approaches
Baseline approaches:
Every 500 m apart (primarily at intersections)
Every 1 km apart
Heuristic based on traffic volume:
Deploy RSUs in cells with high traffic volume
Metrics:
Number of RSUs deployed
Message loss due to lack of network connectivity
Message loss due to CPU capacity (cycles) exceeded at
RSUs
20. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
20/28
Evaluation: RSUs deployed
Approach Number of RSUs
RSU-OPT (α = γ = 100%) 105
Traffic volume 84
Every 500 m 91
Every 1 km 42
21. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
21/28
Evaluation: message loss
RSU-OPT Traffic
Volume
Uniform
(500m)
Uniform
(1km)
0
5
10
15
20
25
30
35
Messageloss(%)
moderate traffic
heavy traffic
RSU-OPT results in lowest
message loss
Uniform (500 m): almost
double the loss despite
having 91 RSUs
Traffic volume heuristic:
low message loss with only
84 RSUs
22. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
22/28
Evaluation: CPU capacity
RSU-OPT Traffic
Volume
Uniform
(500m)
Uniform
(1km)
0
2
4
6
8
10
Messageloss(%)
moderate traffic
heavy traffic
RSU-OPT: lowest
message loss, lowest
spread
Traffic volume heuristic:
higher message loss
despite placing RSUs in
areas with high traffic
volume
23. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
23/28
Evaluation: RSU-OPT settings
γ (%)
60
65
70
75
80
85
90
95
100
α (%)
60
65
70
75
80
85
90
95
100
NumberofRSUs
30
40
50
60
70
80
90
100
110
Reducing network coverage (γ) and computational
demand (α) by 5% results in only 85 RSUs
Target network coverage (γ) is more stringent constraint
24. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
24/28
Evaluation: RSU-OPT settings
γ =95
α =95
85 RSUs
γ =90
α =90
73 RSUs
γ =90
α =95
74 RSUs
γ =90
α =97
74 RSUs
γ =90
α =100
104 RSUs
0
2
4
6
8
10
12
Messageloss(%)
γ=100,
α=100
Traffic
Volume
Uniform
(500m)
γ =95
α =95
85 RSUs
γ =90
α =90
73 RSUs
γ =90
α =95
74 RSUs
γ =90
α =97
74 RSUs
γ =90
α =100
104 RSUs
0.0
0.2
0.4
0.6
0.8
1.0
1.2
Messageloss(%)
γ=100,
α=100
Traffic
Volume
Uniform
(500m)
RSU-OPT with α = γ = 95: similar to traffic volume heuristic
RSU-OPT with γ = 90 and 74 RSUs: lower message loss than
Uniform (500 m) with 91 RSUs
26. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
26/28
Summary and Future Work
Proposed a MILP to efficiently deploy edge computing
devices in a dense urban scenario
What next?
Examine effect of different thresholds for delay
Extend static deployment to a dynamic scenario
Different wireless communication technologies
28. Efficient placement of edge computing devices for vehicular applications
G. Premsankar, B. Ghaddar, M. Di Francesco, R. Verago
NOMS 2018
28/28
Overview of evaluation