Professional Documents
Culture Documents
net/publication/269328393
CITATIONS READS
12 287
7 authors, including:
Some of the authors of this publication are also working on these related projects:
Network-as-a-Service and Network Virtualization for Unification of Networking and Cloud Computing View project
All content following this page was uploaded by Qiang Duan on 06 March 2017.
Abstract—The large amount of data collected in the Internet the IoT for service provisioning. On the other hand, parts
of Things (IoT) need to be transmitted to servers for processing of IoT, especially in sensor networks and wireless access
in order to provide various services. Due to the limited amount of networks, often have limited network bandwidth and energy
resources in IoT, including network bandwidth, node processing
abilities, and server capacities, congestion control in IoT plays a [3], [4]. Servers in the processing platform also only have finite
crucial role for meeting service performance requirements. In this capacities. Therefore, congestion control plays an important
paper, we propose a model for congestion control in IoT with an role in IoT for satisfying service performance requirements.
improved Random Early Discard (IRED) algorithm. We employ Modeling and analyzing congestion control for service
queueing theory to analyze the performance of the proposed
control mechanism. We also conduct extensive simulations to access in IoT are challenging issues for accommodating the
evaluate performance of the proposed control and compare increasingly diverse range of traffic that must be confronted by
it with regular RED algorithm. Our analysis and simulation the users. Previous work mainly aimed at reliability and cost
results show that the proposed control achieves comparable delay of IoT service, without consideration for IoT service access.
performance and better throughput performance compared to The management of IoT service access is vital to obtain a full
standard RED. The simple control mechanism of IRED makes
it more suitable to be implemented in IoT. understanding of system performance. In this paper, we focus
on the modeling and analysis of IoT service control and make
I. I NTRODUCTION the following contributions.
Last years have witnessed the dawn of a new era of Internet • We present a hierarchical system framework, by which
of Things (IoT). IoT is a new networking paradigm that is we model the congestion control process in IoT.
envisioned to interconnect trillions of smart devices or things • We propose an improved Random Early Drop (IRED)
providing and consuming information on the network [1]. algorithm based on the model. Our algorithm employs
In IoT physical things/devices are equipped with different the instantaneous queue length to calculate drop rate of
kinds of sensors and actuators and connected to the Internet the system, which guarantees the real time response of
via heterogeneous access networks enabled by technologies IoT services.
such as embedded sensing and actuating, radio frequency • We employ queueing theory to develop an analysis tech-
identification (RFID), wireless sensor networks, real-time and nique for evaluating performance of the proposed control.
semantic web services, etc [2]. Thanks to these underlying We then obtain the performance metrics such as average
technologies, IoT is opening tremendous opportunities for a queue length, average delay and total loss probability.
large number of novel applications. • We conduct comparison between IRED and standard
RED through simulations, which indicate that IRED can
With the vast number of things/objects and sensors/actuators
achieve better throughput performance without sacrificing
connected to the Internet, a huge amount of data can be
delay performance. The simple control mechanism of
collected through IoT from the physical world. These data
IRED makes it more suitable to be implemented in IoT.
must be transmitted to servers for processing in order to offer
various IoT services. In order to process the huge amount The rest of this paper is organized as follows: Section II
of data, a data processing platform should be built over the briefly reviews the related works. In section III, the hierarchi-
Internet. Such platform is usually enabled by the service cal system framework for IoT, the congestion control model
technologies that acts as the “brain” of IoT, which refines and a new congestion control algorithm are given in detail.
the raw data to form IoT services provided to users. A large Section IV analyzes the performance of IoT service. Section
amount of data are collected, transmitted, and processed in V shows simulation results and compares the performance of
435
IEEE ICC 2014 - Ad-hoc and Sensor Networking Symposium
436
IEEE ICC 2014 - Ad-hoc and Sensor Networking Symposium
O O O O O O (1 f1 (k )) O (1 f1 (k )) O (1 max p ) O (1 max p ) K
−1
thmax
ρk
0 1 n thmin thmax K1 K M= (1−maxp )(k−thmax ) (1−f (i)) .
k=thmax +1 thmin
n(k−n) n!
P 2P nP nP nP nP nP nP nP nP
than n. When the number of packets is larger than n, the The average delay can be obtained from Little’s law as
arriving packets will wait for the service in the buffer based
L
on the congestion control algorithm. When the buffer is full, W = (5)
the system will drop the arriving packets at once. When the S
system gets into the steady state, the state transition diagram where S is the mean throughput which is given by S =
of congestion control with IRED is shown in Fig. 3.
n−1
K
pk kμ + pk nμ.
In order to find the steady state probability distribution, k=1 n
we assume the exponential distribution for arrival rate and The total loss probability is the sum of each loss probability
exponential service rate. The symbols of arrival rate is set to in the queue, which is given by
λ, similarly, the service rate is set to μ. Then the quantitative
expressions are given as below. D = D1 + D2 + D3 (6)
−1
thmax
K−1
⎧ where D1 = pk f1 (k), D2 = pk maxp and
⎪
⎪ λp0 = μp1 , k=0 k=thmin k=thmax
⎪
⎪ D 3 = pK .
⎪
⎪ λp k−1 + (k + 1)μp k+1 = (λ + kμ)p k , 0<k≤n
⎪
⎪
⎪
⎪ λp k−1 + nμp k+1 = (λ + nμ)p k , n ≤ k < th min V. S IMULATION R ESULTS
⎨
λ(1 − f1 (k))pk−1 + nμpk+1 =
In this section, simulation results are given to illustrate
⎪
⎪ (λ(1 − f1 (k)) + nμ)pk , thmin ≤ k < thmax
⎪
⎪ the performance of the congestion control algorithm. Our
⎪
⎪ λ(1 − max p )p k−1 + nμp k+1 =
⎪
⎪ experiments are performed in the OMNet++ simulator with
⎪
⎪ (λ(1 − maxp )) + nμ)pk , thmax ≤ k < K
⎩ IDE toolkit, which is an object-oriented C++ component-based
λ(1 − maxp )pK+1 = nμpK , k=K
(2) modular and discrete event network simulation framework
(k−thmin )2 primarily for building network simulators [16]. Within the
where f1 (k) = (th −th )2 maxp .
max min OMNet++ IDE, each service request is treated as an event.
Let ρ = λ/μ, then
⎧ A. Simulation Setup
⎪
⎪ρp0 , k=1
⎪
⎪ In our simulation, λ and μ are denoted as the arrival rate and
⎪
⎪ ρ k
⎪
⎪ k! p0 , 1<k≤n service rate, and they are given according to the arrival packets
⎪
⎨ ρk p ,
n(k−n) n! 0
n < k ≤ thmin per second. Network topology can be described in a file. We
pk = k−1 ρk set the thresholds of minimum queue length thmin = 10 and
⎪
⎪ i=thmin (1 − f (i)) n(k−n) n! p0 , thmin < k ≤ thmax
⎪
⎪ maximum queue length thmax = 30, buffer size K = 40, and
⎪
⎪ (1 − max ) (k−thmax )
⎪
⎪ p
maxp = 0.1. All these parameter settings are derived from
⎪
⎩thmax −1 (1 − f (i)) ρk p , th
i=thmin n(k−n) n! 0 max < k ≤ K [17].
K (3)
From i=0 pk = 1, we can get B. Results
In our simulation experiments, we considered congestion
1
p0 = control on traffic for accessing 5 servers and set service rate
G+H +W +M
of each server as μ = 0.02. The minimum threshold and
where maximum threshold are set as 10 and 30 respectively, and the
n
ρk total buffer size is 40. We measured the average queue length,
G= ,
k! average packet delay for service access, and packet loss rate
k=0
under various arrival rates. The obtained results are plotted in
th
min
ρk Figs. 4-6 respectively.
H= , Fig. 4 gives the average queue length with various arrival
k=n+1
n(k−n) n!
traffic load under congestion control of regular RED and
th
k−1
improved RED. The figure shows that the average queue length
max
ρk
W = (1 − f (i)) , for both cases increase with traffic load. The average queue
k=thmin +1 thmin
n(k−n) n! length varies in the range of 0 to 24, which is because the
437
IEEE ICC 2014 - Ad-hoc and Sensor Networking Symposium
500
25 RED RED
IRED IRED
400
20
Average Queue Length
300
Packets Dropped
15
200
10
100
5
0 0
0.6 0.8 1.0 1.2 1.4 1.6 1.8 0.6 0.8 1.0 1.2 1.4 1.6 1.8
Traffice Load Traffic Load
Fig. 4. Average queue length vs. arrival rate Fig. 6. Packet loss vs. arrival rate
40
37
Average Delay
36
300
35
200
34
33
100
32
0 31
0.6 0.8 1.0 1.2 1.4 1.6 1.8 35 36 37 38 39 40
Traffic Load th2-th1
maximum threshold is set as 30 in the queue. With increasing significantly with the arrival rate. Comparison between the
in load, both average queue curves first raise significantly and two curves in the figure indicates that IRED has a lower
then tend to approach flat even under relatively heavy traffic packet drop rate than RED under relatively heavy traffic load.
load. Fig. 5 shows the average packet delay under various This implies a better throughput performance achieved by the
traffic load with RED and IRED control. We noticed that both proposed IRED congestion control.
delay curves increase with traffic load by following a similar Then we varied the maximum and minimum thresholds
patten as Fig. 4. These two figures indicate that both queue and tested the impact of threshold difference on average
length and average delay are upper limited even under heavy queue length, average delay, and packet drop performance. The
traffic control due to the effect of congestion control. obtained results are given in Figs. 7, 8 and 9. From Figs. 7 and
Comparing queue length and average delay performance of 8 we can see that both average queue length and average delay
RED and IRED shown in Figs. 4 and 5, we can see that RED increase obviously with increase in maximum threshold. Fig. 9
and IRED achieve very close average queue length as RED and shows that the number of dropped packets decrease when the
slight better packet delay performance under most traffic load. gap between the maximum and minimum thresholds. This is
This shows that our proposed IRED can achieve comparable because a large maximum threshold makes it more likely to
delay performance as standard RED with a simplified control accept an arrival packet into the queue; therefore, reduce the
mechanism (simpler algorithm for evaluating packet drop number of dropped packet and increase the average queue
probability) that is easier to be implemented in IoT. length and packet delay.
Fig. 6 gives the number of dropped packets for RED and In summary, The above reported simulation results indicate
IRED under various traffic load. The figure shows that for that the proposed IRED congestion control can achieve a
both IRED and RED the number of dropped packets increase better throughput performance than the regular RED without
438
IEEE ICC 2014 - Ad-hoc and Sensor Networking Symposium
740
stitutions of Higher Education in Chongqing (Grant No.
KJTD201310), Natural Science Foundation of Chongqing,
730
(Grant No. cstc2013jcyjA40026), Scientific and Technological
720 Research Program of Chongqing Municipal Education Com-
710 mission (Grant No. KJ130523), and CQUPT Research Fund
700
for Young Scholars (Grant No. A2012-79).
690 R EFERENCES
35 36 37 38 39 40 [1] L. Atzori, A. Iera, and G. Morabito, “The internet of things: A survey,”
th2-th1 Computer Networks, vol. 54, no. 15, pp. 2787 – 2805, Oct. 2010.
[2] X. Cao, J. Chen, Y. Zhang, and Y. Sun, “Development of an integrated
wireless sensor network micro-environmental monitoring system,” Isa
Fig. 8. Average delay vs. difference between thresholds
Transactions, vol. 47, pp. 247–255, 2008.
[3] J. Chen, S. He, Y. Sun, and P. Thulasiraman, “Optimal flow control
360 for utility-lifetime tradeoff in wireless sensor networks,” Computer
RED Networks, vol. 53, pp. 3031–3041, 2009.
340 [4] J. Chen, W. Xu, S. He, Y. Sun, P. Thulasiraman, and X. Shen,
IRED
“Utility-based asynchronous flow control algorithm for wireless sensor
320 networks,” IEEE Journal on Selected Areas in Communications, vol. 28,
pp. 1116–1126, 2010.
300 [5] L. Li, Z. Jin, G. Li, L. Zheng, and Q. Wei, “Modeling and analyzing
Packets Dropped