International Journal of Computer Science and Applications,

Technomathematics Research Foundation

Vol. 13, No. 1, pp. 31 – 41, 2016

A COGNITIVE NETWORK BASED ADAPTIVE LOAD BALANCING

ALGORITHM FOR EMERGING TECHNOLOGY APPLICATIONS*

Kang Lu†
Science and Technology on Space Physics Laboratory
Beijing, 100076, China‡
[email protected]§

Ting Xing
IDC(Beijing) Co,Ltd,
Beijing, 100044, China
[email protected]

In the era of cloud computing and big data, the demand for real-time data processing and availability
poses higher requirements for network load balancing. Cognitive network has unique self-learning
and re-configuration abilities that can improve the effectiveness of load balancing. Based on the
existing traffic scheduling algorithms, this article will discuss the possibility of improving weighted
least-connection scheduling algorithm by leveraging cognitive network. A dynamic load balancing
algorithm (NNPMA,Neural Network Prediction Model Algorithm) will be developed on the basis of
traffic prediction model. NNPMA can enable least load scheduling in real time for service request
from node and configure available idle resources upfront to ensure compliance with QoS. This
traffic scheduling algorithm will be simulated with OPNET and it will be applied to cloud
computing architecture. The test results indicate that this algorithm can achieve loading performance
better than the unimproved weighted least-connection scheduling algorithm without significantly
increasing network overhead.

Keywords: Cloud Computing; Big Data; Cognitive Network; Load Balancing.

1. INTRODUCTION
As cloud computing embarks on the stage of real application, hybrid cloud service
becomes the mainstream model provided from cloud services at present stage, and big
data analysis emerges to be a hotspot of cloud computing application. Versatile and
massive data is generated explosively, data is growing at an unprecedented pace around
the world; network access users, traffic and access requests increase by folds, all
demanding higher network and data processing quality. In such context, a data center
can't deploy its servers in the traditional pattern of single point integration, but leverages
highly-efficient innovative service system architecture to realize more flexible cluster
deployment and more reasonable resource configuration for building a resource server

31
32 Lu Kang and Xing Ting

pool. However, a heterogeneous cloud computing application platform will have resource
server instances in various sizes, with an excessive load may create unbalanced loading at
server side, so load balancing plays a crucial part in ensuring the efficient operation of
cloud computing system architecture. Working on the design of cloud computing system
architecture, the cognitive network adaptive load balancing algorithm can flexibly turn
on/off cloud service resources and adaptively configure request for network service to
optimal server instance, ensuring the uninterrupted services.
Currently load balancing has not been integrated into cloud computing platform, but
works as a rather independent module. At present factors like distribution of terminals,
random mobility and various QoS requirements for user services may cause uneven
distribution of network traffic, leading to local node overloading and congestion at nodes
with heavy load that increase packets lost rate and service latency, while idle resources at
nodes with light load are in poor utilization. Cognitive network load balancing offers an
innovative solution to this scenario.
LVS cluster architecture implements load balancing by modifying IP to schedule
resources, and as the virtual instance resources on cloud computing service platform all
have their own ID number and IP addresses, it is very easy to apply LVS cluster
technology to enable adaptive configuration of cloud computing resources. But common
scheduling algorithms of LVS can't be used in a scenario in which resources are changing
dynamically. The NNPMA algorithm proposed by this article enables a load balancing
and scheduling scheme suitable for use in cloud computing system architecture. While
ensuring the compliance with user QoS requirements, it can configure upfront available
network resources in accordance with load variation to schedule network traffic in real
time and dynamically so that traffic can be distributed evenly across the entire network to
reduce network congestion, offering a more effective solution for today's big data
application processing and the extensive use of cloud computing.

2. CURRENT STATE OF RESEARCH

Scheduling algorithm is at the center of implementing load balancing. The traffic
scheduling algorithms present in use have time latency, for they schedule and distribute
traffic to balance load by following the results of current network state and the analysis of
captured parameters [Kong et al,(2007)][Luo et al,(2008)][Li et al,(2005)][2-4].
Apparently the traffic scheduling in current network state has latency, while due to its
upfront learning capability, cognitive network can essentially understand the upcoming
traffic and evenly distribute network traffic to each server before the arrival of network
service request, so as to maintain network load in balance. Load balancing is mainly
designed to complete tasks like: overcome network congestion and provide service at
proximity to realize geographical-location irrelevance; provide users better access
qualities through accelerating the response speed of services; and improve the utilization
of server and other resources to avoid single point failure happening at mission-critical
part of the network. Scheduling algorithms used in today's network mainly include [Kong
et al,(2007)][Luo et al,(2008)]: round-robin scheduling, weighted round-robin
 A Cognitive Network Based Adaptive Load Balancing Algorithm for Emerging Technology Applications 33

scheduling, least-connection scheduling, weighted least-connection scheduling,

destination address HASH scheduling and source address HASH scheduling. These
algorithms can also be divided into static traffic scheduling and dynamic traffic
scheduling, with the expansion of network and users, static traffic scheduling is no longer
able to meet the requirements of load balancing, so a dynamic and adaptive traffic
scheduling algorithm shall be developed to more closely meet the requirements from
today's network applications.
The above scheduling algorithms play a certain role in solving server load balancing.
However, with the growing size of the network and increase in the access requests,
deficiencies exist in the scheduling performance of these scheduling algorithms [Li et
al,(2005)][Liang et al,(2009)][K et al,(2004)][Yan et al,(2005]. They do not fully
consider the load status of the server, and the round-robin scheduling algorithm and
weighted round-robin scheduling algorithm are the stateless scheduling algorithm, and
they are static scheduling [Li et al,(2005)]; the destination address hashing and source
address hashing algorithm is a static mapping algorithm; for the least connection
scheduling and weighted least connection scheduling algorithm [Li et al,(2005)],
although they take into account the state of number of the server’s connection requests
and they are the dynamic scheduling, the server load is not just the number of connection
requests, and other status information also needs to be considered. On the other hand,
these scheduling algorithms cannot give feedback effectively based on real-time
comprehensive state of the server [Hu et al,(2009)][Chen et al,(2006)][Wang et
al,(2006)], and schedule the load request to the appropriate server. Most of the existing
studies aim at the general packet data network, and conduct traffic scheduling by use of
global status information obtained at a certain moment. Since it is difficult to obtain
accurate global state information, and the performed scheduling cannot be adjusted in
real time, thus the load balancing effect is limited and lags significantly [Wang et
al,(2006)] [Chen et al,(2008)][FU et al,(2006)][Zhang et al,(2007)]. The network state
learning capability of cognitive network enables the router to have environmental
learning and adaptive processing mechanism, effectively solving the scheduling lag issue.
Weighted least connection scheduling algorithm is the most efficient traffic scheduling
algorithm among the basic dynamic scheduling algorithms [Zhang et al,(2007)], but it
only takes into account the number of connection requests to the server, and CPU load,
network traffic and memory usage are also important parameters affecting the load
balancing.
In addition, the network traffic in real world exhibits features like strong nonlinearity and
multi-scale, such as fractal, long-range correlation, self-similarity and abruptness; the
cognition of the cognitive network is embodied in its ability to predict network traffic.
Because traditional network traffic model can only handle steady process and special
unsteady process, more discrepancy happens when describing network traffic behaviors.
To address the multi-scale and nonlinearity of computer network traffic system, a new
combined model that integrates the multi-resolution analysis of wavelet transformation
and the nonlinear approximation of neural network needs to be built to predict network
traffic.
34 Lu Kang and Xing Ting

Fig. 1.Improved Wavelet Neural Network Prediction Combined Model Algorithm Scheme

The basic idea of this paper is to improve wavelet neural network prediction combined
model created by integrating wavelet neural network model; improve weighted least-
connection scheduling algorithm; propose NNPMA and apply it to cloud computing
system architecture.

3. NNPMA ALGORITHM

3.1. Algorithm Scheme

The improved wavelet neural network prediction combined model algorithm scheme is
illustrated in Fig 1. It can predict the upcoming network traffic from the last actual
network traffic flowing through prediction model and then accordingly configure network
resources.
Though weighted least-connection scheduling algorithm is a widely used load
balancing algorithm with nice effects [Ron and Ariel, (2007)], it has included the factors
causing imbalance of server performance into consideration and is unable to aggregate all
server parameters to understand its performance. To tackle this drawback, this paper has
improved this algorithm by designing an adaptive scheduling algorithm bases upon
cognitive network, i.e. the NNPMA.
Cognitive network can learn network traffic status parameters by itself and understand
server performance from these parameters. After fully understanding server performance,
new network service requests together with server management thresholds will be brought
under weight analysis, to derive new parameters about traffic flow. With these new
parameters, new requests can be scheduled in such a way that new request will not be
scheduled to a server with least load but low performance, thus avoiding causing new load
imbalance, which is the case with least-connection scheduling algorithm.
With this thought in mind, the basic flow chart is illustrated as Fig 2:
 A Cognitive Network Based Adaptive Load Balancing Algorithm for Emerging Technology Applications 35

Fig. 2. Core Algorithm Basic Flow Chart

In the flow of this algorithm, the first step is to predict the upcoming network traffic
according to the basic traffic; then derive the status parameters of upcoming network
traffic from the traffic output predicted by traffic prediction model; after fully considering
the impact patterns of various services on network traffic, a specialized network software
MRTG will be used to capture the status parameters of server; calculate the weighted load
value of each server, and record the largest and the least weighted value, and the traffic
between paths can then be dynamically adjusted with network resource adaptive
distribution approach. For service with connection established as well as newly connected,
their types and resource requirements will be compared comprehensively and the
comparison results will be reported to the router node at the forward path. In other words,
the largest weighted value of servers will be compared with their management thresholds.
Corresponding servers will stop responding for a certain period of time decided by their
working state, and the latest load request will be scheduled to the server with least load.
After this, these nodes will redirect those packets of existing service flows according to
assessment results to enable real-time and dynamic scheduling of service flows in
network.

3.2. Algorithm Implementation

Parameter description: There are m servers, and j ranges from 1 to 3, representing CPU
load, network traffic and memory usage respectively.
ri [ j ](1  i  m,1  j  3)
(1)
W j (1  j  3 )
(2)
Eq. (1) represents the CPU load, network traffic, memory usage load parameters of the ith
server, while Eq. (2)represents the weighted value of server CPU load, network traffic, and
memory usage. Load for each server i is Eq. (3)
Load[i]   ri [ j ]  W j
(3)
Proxy server, according to the load of each server, schedules the request by the scheduling
algorithm to the appropriate server.
Algorithm is described as follows:
MaxLoadValue=0;
For (int i =1; i <= m; i++)
{
For (int j=1;j<=3;j++)
36 Lu Kang and Xing Ting

Load[i]=Load[i]+r[j]?Wj
If (MaxLoadValue< Load[i])
{
MaxLoadValue = Load[i];
Return MinLoad[i]
}
}
If (MaxLoadValue< predetermined threshold)
Suspend server [i] for a certain time;
else
Obtain performance parameters of several sets of servers and calculate the weighted
value.
Return the current minimum load server algorithm pseudo-code:
If W(Smin)>0
{for (int i=m+1;i<n;i++)
{if (W(Si)<=0)
Continue
If (Load(Si)<Load(Smin))
M=i
}
Return Smin
}

4. SIMULATION IMPLEMENTATION

4.1. Create Simulation Model

OPNET [Li and Ye, (2006)] is a universal simulation and modeling tool for
communications network and facilities. Its object-oriented modeling method and graphical
editor can truly reflect the actual network and structure of various network components,
and the system can be mapped directly into the model. Its flexibility almost enable it to
support all types of network and network technologies, and to be widely used as a
decision-support tool for network to conduct detailed insight and analysis on the existing
and in-design network, system or process performance and behavior.
Forming in common network structure, the network simulation model includes three
parts as input, processing and output. The input generates network data using an
autoregressive model based on Poisson distribution and Bernoulli distribution. The process
model in router captures network parameters and server performance metrics, calculates
weighted load value and outputs load scheduling instructions to egress node. As shown in
Fig 3.
 A Cognitive Network Based Adaptive Load Balancing Algorithm for Emerging Technology Applications 37

Fig. 3. Schematic Drawing of Simulation Model

The basic adaptive algorithm designed in Section 2.1 is improved and included into the
simulation model to set up process model, node model and engineering scenario.
A network simulation scenario model is built, in which three clients are configured to
transmit and receive data from three servers. Then the weighted least-connection
scheduling algorithm is compared with improved NNPAM algorithm, where certain basic
parameters are set to reflect common network data. For example, the process model is set
as “Acb_fifo” (FIFO); the average interval between each arrival of packets as 1.0s; the
size of data packet as 1,000 to 9,000b/p; the transmission interval as 1s, and the simulation
duration as 30 min.

4.2. Simulation Result Analysis

When the simulation ends, results of parameters are gathered and then analyzed.
As Fig 4 Fig 5 shows, the traffic curve generated by improved prediction model well
fits the actual traffic curve, enabling to predict upcoming network traffic with high
accuracy. Particularly in the late phase of simulation, the traffic results of improved
prediction model are even more close to actual traffic. So the network traffic generated by
improved combined prediction model can be safely used to extract parameters about the
upcoming network traffic.
38 Lu Kang and Xing Ting

Fig. 4. Predicting Curve Generated by Improved Prediction Model and Actual Traffic Value

Fig. 5. Predicting Curve Generated by Unimproved Prediction Model and Actual Traffic Value

The load rendering of each server is shown in Fig 5 Fig6. The horizontal axis
represents 30min simulation time, and the vertical axis represents the loads of different
servers.
 A Cognitive Network Based Adaptive Load Balancing Algorithm for Emerging Technology Applications 39

Fig. 6. Load Rendering of Minimum Traffic Scheduling Server

Fig. 7. Load Rendering of Adaptive Minimum Traffic Scheduling Algorithm Server

The load situations of each serve under the minimum traffic scheduling algorithm are
shown in Fig 6, and Fig 7 shows the load situations of each serve under the improved
adaptive minimum traffic scheduling algorithm; the loads of the three servers shown in Fig
6 are not balanced enough, and the third server takes more network requests. In Fig 7, the
40 Lu Kang and Xing Ting

loads of the three servers are relatively balanced, and they take the network requests in a
relatively average manner. After using improved NNPMA algorithm, the load balancing
for each server is better than using the minimum traffic scheduling algorithm.

5. Application of NNPMA Algorithm in Cloud Computing Architecture

Cloud computing platform provides flexible server resource instance for the load
balancer, after significant changes of the user request access volume and adaptive
adjustment of resources, conduct the adaptive adjustment to the load request scheduling,
achieving the consistency between the request load adaptive scheduling and resource
supply. Adaptive load balancing scheduling of the server instance resource is achieved via
the appropriate deployment of LVS cluster system scheduler. Save the NNPMA algorithm
in LVS scheduling algorithm module, the load status judgment and load forwarding are
achieved through a variety of linked lists; the administrator calls IPVS-related functions
through the configuration of management module interaction, achieving load balancing
resource scheduling of LVS cluster system. The implementation of this application is
shown in Fig 8.

Fig. 8. Implementation of NNPMA Algorithm in Cloud Computing Application

6. Conclusions
In a context of extensive use of emerging technologies today, by applying NNPMA to
cloud computing architecture and building model with simulation software OPNET to
simulate the algorithm. Then compare and analyze the results of NNPMA with
unimproved least-connection scheduling algorithm, it can be concluded that the dynamic
and adaptive traffic scheduling algorithm can enable servers to evenly provide network
services and better balance load, while minimizing impact to overall network performance,
making it a more optimal solution for current cloud computing application architecture.
 A Cognitive Network Based Adaptive Load Balancing Algorithm for Emerging Technology Applications 41

References
Chen Jianxun, Zhang Yong, et al. A Load Balancing Algorithm for MPLS Traffic Engineering
Field. Computer and Digital Engineering, no. 6, 2006, pp. 39-41.
Chen Yijiao, Lu Xicheng, et al. A Session-Oriented Adaptive Load Balancing Algorithm. Journal
of Software, no. 19, 2008, pp. 1828-1836.
FU Guowei, Liu Xinsong, et al. A Net Load Scheduling Algorithm Based on Feedback. Journal of
Chengdu University of Information Technology, no. 12, 2006, pp. 779-781.
Hu Anbo, Su Jinshu, Chen Shuhui, et al. Research on Application-Level Load Balancing
Technology. Computer Engineering and Applications, no. 45, 2009, pp. 84-86.
K. Gopalan, Tzi-cker Chiueh, Yow-Jian Lin, “Load balancing routing with bandwidth-delay
guarantees”, IEEE Communications Magazine, 2004, pp. 108-113.
Kong Dawei, Li Dandan, et al. Summarization of Load Balancing Algorithms for Network
Processors. Automation Technology and Applications, no. 26, 2007, pp. 45-48.
Li Daisong, Liu Yong, et al. Study and Implementation of One Kind of Load Balancing
Algorithm. Journal of Shenyang Ligong University, no. 9, 2005, pp. 47-48.
Li Xin, Ye Ming. Network modeling and simulation for OPNET Modeler. Xi'an: Xi'an University
of Electronic Science and Technology Press, 2006.
Liang Benlai, Qin Yong, Chen Li, Song Fei, et al. DBCTIA Algorithm Based on Binary Object
Optimization for Load Balancing of Multiple Links. Computer Applications, no. 3, 2009, pp.
655-664.
Luo Yongjun, Li xiaoyue, et al. Summarization of the Load-balancing Algorithm. Sci-Tech
Information Development & Economy, no. 23, 2008, pp. 134-136.
Ron Banner and Ariel Orda, “Muttipath routing algorithms for congestion minimization,”
IEEE/ACM Transactions on Networking, vol. 15, no. 2, 2007, pp. 413-424.
Song Huili, et al. Research on Key Technologies of Cognitive Radio Networks. Mobile
Communications, no. 4, 2008, pp. 75-77.
Wang Yue, Cai Wandong, et al. An Adaptive Dynamic Load Balancing Algorithm. Computer
Engineering and Applications, no. 21, 2006, pp. 121-123.
Yan Shi, Zengji Liu, Zhiliang Qiu and Min Sheng. “Load Balance Based Network Bandwidth
Allocation for Delay Sensitive Services” Proc. 19th International Conference on Advanced
Information Networking and Applications (AINA’05), 2005.
Zhang Weiwen, Wu Guoxin, et al. Research on Dynamic Load Balance Algorithm Based on
Structured P2P Networks. Computer Engineering and Design, no. 9, 2007, pp. 4152-4168.