Content uploaded by Kolichala Rajashekar
Author content
All content in this area was uploaded by Kolichala Rajashekar on Aug 20, 2023
Content may be subject to copyright.
Topology-Aware Cluster Configuration for Real-time Multi-access
Edge Computing
Kolichala Rajashekar
kolichalar@iitbhilai.ac.in
Indian Institute of Technology Bhilai
Raipur, India
Sushanta Karmakar
sushantak@iitbhilai.ac.in
Indian Institute of Technology Guwahati
Guwahati, India
Souradyuti Paul
souradyuti@iitbhilai.ac.in
Indian Institute of Technology Bhilai
Raipur, India
Subhajit Sidhanta
subhajit@iitbhilai.ac.in
Indian Institute of Technology Bhilai
Raipur, India
ABSTRACT
We consider data-intensive real-time systems, such as mission-
critical data-intensive applications such as forest re detection,
medical emergency services, oil pipeline monitoring, etc., which
demand relatively low response time in processing data from IoT
(Internet of Things) devices. Usually, in such cases, the edge com-
puting paradigm is leveraged to drastically reduce the processing
delay of such applications by performing the computations on edge
devices placed closer to the data sources, i.e., the IoT devices. How-
ever, most edge devices, such as cellular phones, tablets, and UAVs
(Unmanned Aerial Vehicles), are mobile in nature. Hence, the clus-
ter conguration must be dynamically adapted with respect to the
changing network topology of the edge cluster such that the ob-
served overall communication delay incurred by the edge devices
in processing the data from the IoT devices is minimized. To that
end, we propose Deep Reinforcement Learning-based intelligent
assignment of IoT devices to non-stationary edge devices such that
the communication delay is minimized and none of the edge devices
is overloaded. We demonstrate, with some preliminary results, that
our algorithm outperforms the state-of-the-art.
KEYWORDS
IoT, edge computing, reinforcement learning
ACM Reference Format:
Kolichala Rajashekar, Sushanta Karmakar, Souradyuti Paul, and Subhajit
Sidhanta. 2018. Topology-Aware Cluster Conguration for Real-time Multi-
access Edge Computing. In Proceedings of ACM Conference (Conference’17).
ACM, New York, NY, USA, 2 pages. https://doi.org/XXXXXXX.XXXXXXX
1 INTRODUCTION
Though the edge computing paradigm eectively reduces the re-
sponse time by processing the data from the IoT devices closer
to the data sources, maintaining a low response with increasing
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specic permission and/or a
fee. Request permissions from permissions@acm.org.
Conference’17, July 2017, Washington, DC, USA
©2018 Association for Computing Machinery.
ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. . . $15.00
https://doi.org/XXXXXXX.XXXXXXX
number of IoT devices using resource-constrained edge devices is
a challenging task. Additionally, the mobility of IoT and edge de-
vices introduce more challenges to the above problem such as links
breakdown, reconguration overhead, churning etc.[2]. Therefore,
in such a dynamic edge cluster, the IoT devices must be assigned to
non-stationary edge devices while taking into account the dynamic
changes in the topology of the edge cluster and the relative change
in the distance of the IoT devices to the corresponding edge de-
vices. Even with the stationary IoT and edge devices, the problem
is Np-Hard as this problem is similar to a well-known NP-Hard
problem, i.e., Generalised Assignment Problem (GAP) [1]. Hence,
obtaining an optimal solution, in this case, is theoretically infeasible.
Therefore, we apply heuristics approaches to get a near-optimal
solution depending on the given deployment scenario. In this pa-
per, we present a novel approach for managing a non-stationary
edge cluster such that the communication delay in processing data
from IoT devices is minimized. Specically, we demonstrate that
Deep Reinforcement Learning (DRL) based approaches are able
to produce a near-optimal assignment of IoT devices to edge de-
vices, maintaining a reduced communication delay with changing
topology.
2 PROBLEM DEFINITION
Let there be
𝑚
IoT and
𝑛
mobile edge devices in the network. Let
𝐿(𝑖, 𝑗 )
𝑡
denote a pair of locations of the
𝑖
th IoT and the
𝑗
th edge device
at time
𝑡
. The mobility function
M
takes the previous location-pair
𝐿(𝑖, 𝑗 )
𝑡′
as input and outputs the current location-pair
𝐿(𝑖, 𝑗 )
𝑡
where
𝑡>𝑡′
. We dene the communication delay
𝐶(𝑖, 𝑗 )
𝑡
recursively using
the mobility function Mas follows.
𝐶(𝑖, 𝑗 )
𝑡=𝑓(𝐿(𝑖, 𝑗 )
𝑡, 𝐿 (𝑖,𝑗 )
𝑡′)
=𝑓(M (𝐿(𝑖, 𝑗 )
𝑡′, 𝑡, 𝑡 ′), 𝐿 (𝑖,𝑗 )
𝑡′)
=𝑔(𝐿(𝑖, 𝑗 )
𝑡′, 𝑡, 𝑡 ′)
An assignment
𝐴𝑡
is a binary matrix, where an element of it
𝐴(𝑖, 𝑗 )
𝑡is dened as follows:
𝐴(𝑖, 𝑗 )
𝑡=(1,if 𝑖th IoT device is connected to the 𝑗th edge device at time 𝑡
0,otherwise.
The set of all possible assignments is denoted by
A={𝐴𝑡|𝑡=
1
,
2
,· · · }
and we note that an IoT device can be connected to only
Conference’17, July 2017, Washington, DC, USA Kolichala Rajashekar, Sushanta Karmakar, Souradyuti Paul, and Subhajit Sidhanta
one edge device at a time; mathematically,
Í𝑚
𝑗=1𝐴(𝑖, 𝑗 )
𝑡=
1
,∀𝑖∈
{
1
,
2
. . . 𝑛},∀𝐴𝑡∈ A
. Therefore, our problem
𝑀𝐴
-
𝐸𝐴𝑃
is as an
optimization problem as dened below:
min max (C𝑡𝐴⊺
𝑡) | 𝐴𝑡∈ A } (1)
Where
C𝑡𝐴⊺
𝑡
produces a matrix of size
𝑚×𝑚
and
max (C𝐴⊺
𝑡)
gives the maximum of all elements of that matrix.
However, the aforementioned optimization problem needs to be
solved under a few constraints as described below.
(1)
The maximum number of IoT devices that can be connected
to an edge device is
T
. (Therefore, given
𝑚
,
𝑛
can be deter-
mined (or vice-versa) from the following equation
𝑚
𝑛≤ T
.)
(2)
The hamming distance between
𝐴𝑡
and
𝐴𝑡′
should be less
than
R
, where
R
is dened as the maximum number of
allowed re-congurations.1
3 HEURISTICS DESIGN
We solve our problem in the following two scenarios where
•
Scenario 1: The IoT devices are stationary and edge devices
are mobile.
•Scenario 2: Both the IoT and edge devices are mobile.
Swarm Intelligence and Machine Learning are extensively used in
managing the resources in edge computing. To that end, we apply
Deep Reinforcement Learning (DRL) and compare with Particle
Swarm Optimization (PSO). The goal of the PSO algorithm is to nd
particle position such that it results in the best evaluation tness
(objective) function. The tness value of an edge device
𝑗
(
𝑓 𝑖𝑡 𝑛𝑒𝑠𝑠𝑗
)
is dened as follows
𝑓 𝑖𝑡 𝑛𝑒𝑠𝑠𝑗=max{𝐶(1, 𝑗 )
𝑡∗𝐴(𝑖, 𝑗 )
𝑡,· · · , 𝐶 (𝑚, 𝑗 )
𝑡∗𝐴(𝑚,𝑗 )
𝑡}(2)
In order to apply DRL, we formulate the dened problem into a
Partially Observable Markov Decision Process (POMDP) framework.
Elements of POMDP are dened as follows
•State Space of an edge device j:
𝑆(𝑗)
𝑡={𝐿(1,𝑗 )
𝑡∗𝐴(1,𝑗 )
𝑡,· · · , 𝐿 (𝑚,𝑗 )
𝑡∗𝐴(𝑚,𝑗 )
𝑡}
•Action Space of an edge device 𝑗:
𝐴𝑗={𝐴(1,𝑗 )
𝑡· · · , 𝐴 (𝑚,𝑗 )
𝑡}
•Reward for an edge device 𝑗:
𝑅𝑗
𝑡=Í𝑚
𝑖=1𝐴(𝑖, 𝑗 )
𝑡+1
max{𝐶(𝑖, 𝑗 )
𝑡∗𝐴(𝑖,𝑗 )
𝑡}∀𝑖∈ {1,· · · , 𝑚}
Considering the above elements of POMDP, we observe that the
state space and the action space are larger. Hence, applying tabular
RL methods is infeasible. Therefore, we use DRL and leverage deep
Q-learning Network (DQN) for our problem.
4 PRELIMINARY RESULTS
We performed simulation using the CRAWDAD dataset traces for
the mobility of IoT devices and random mobility for edge devices.
For realistic analysis, we obtain the communication delays from
the PlanetLab dataset.
Fig.1 represents the communication delay changes as the num-
ber of episodes increases with respect to Scenario 1. Initially, PSO
solution moves towards global optima very fast than DQN. This
1
An IoT device is reassigned to another edge device due to the mobility of the cor-
responding edge device. This re-assignment of IoT device to edge device is called
re-conguration.
105
110
115
120
125
130
135
140
145
150
Communication Delay (ms)
Episodes
DQN
PSO
Figure 1: Changes in communication delay observed during
episodes ( iterations) of Scenario 1
0
50
100
150
200
250
Communication Delay (ms)
Different snapshots of Scenario 2
PSO DQN
Figure 2: Communication Delay observed in dierent snap-
shots
is because of the exploration done by the DQN agent. After cer-
tain episodes, DQN agent learns the environment converges faster
whereas PSO converges very slowly.
Fig.2 shows the observed communication delay. This experiment
was conducted in dierent snapshots, where the DQN agent is
already trained whereas PSO has to run from scratch. Due to the
mobility of the IoT and edge devices, PSO is not able to give better
results than DQN.
REFERENCES
[1]
Juan A Dıaz and Elena Fernández. 2001. A tabu search heuristic for the gener-
alized assignment problem. European Journal of Operational Research, 132, 1,
22–38.
[2]
Zeeshan Hameed Mir, Deepesh Man Shrestha, Geun-Hee Cho, and Young-
Bae Ko. 2006. Mobility-aware distributed topology control for mobile multi-
hop wireless networks. In International Conference on Information Networking.
Springer, 257–266.