Application of Ant Colony Optimization in Cloud

Computing Load Balancing

Wu Zheng-tao

School of information engineering, Shanghai Maritime University, Shanghai, China, 201306

Abstract. In cloud computing, because all the tasks are submitted and requested
by the users, it is the key to improve the quality of service whether all tasks are
scheduled to every host to achieve the shortest completion time for all tasks. In
this paper, the basic ant colony algorithm is modified appropriately to enable task
scheduling and load balancing. The simulation results show that the ant colony
algorithm can effectively obtain the global optimal solution and realize the
shortest task completion time.
Keywords: cloud computing; load balance; Ant Colony Optimization

0 Introduction

Cloud computing is a new business and computing model which is proposed in

recent years. Thousands of hosts are linked together over the network to form an ultra
large pool of resources. Google corporation is the first to use cloud computing service
its scale has reached millions of machines, and other companies such as Microsoft,
Amazon, Aliyun and so on, also reached the scale of several hundred thousand units[1].
Depending on the type of service, cloud computing is divided into three layers, namely,
IaaS (infrastructure as a service), PaaS (platform as a service), and SaaS (software as a
service). No matter which layer, there is the need to balance the task of scheduling to
the host, that is, load balancing. However, due to the large number of computers in the
cloud computing platform, the complexity of the structure and the difference of the
allocation of resources, it is difficult to achieve load balancing for such large-scale,
heterogeneous and non-centralized task scheduling. The unbalanced load will lead to
the decrease of system efficiency and throughput, and the instability of the daily oper-
ation, which will seriously affect the quality of service of cloud computing[2].
Ant Colony Optimization (ACO) is a kind of bionic intelligent algorithm to find
the optimal solution, simulating the foraging behavior of ants. The wisdom of a single
ant is not high, but it achieves an optimal population effect by sharing information
among groups. Ant colony algorithm is a kind of positive feedback algorithm, which
has good robustness and adaptability, and is easy to combine with other algorithms[3].
Therefore, ant colony algorithm can be used in load balancing.
The ant colony system was firstly proposed by Italy scholars Dorigo, Maniezzo
and others in 1990s. When they studied ants foraging, they found that the behavior of
a single ant was relatively simple, but the whole ant colony could show some intelligent
behavior. For example, an ant colony can find the shortest path to a food source in
different environments. This is because ants in the ant colony can transfer information
through some kind of information mechanism. Further studies have shown that ants
release a substance called "pheromones" on their path.
The ants in the colony can apperceive pheromones, they will follow the path wi-
th high concentration pheromone, and every ant will leave the pheromones on the roa
d, which forms a kind of positive feekback mechanism. Thus, after a period of time,
the whole ant colony will arrive at the food source along the shortest path.
The basic idea of applying ant colony algorithm to solve the optimization problem
is to use the path of the ant to represent the feasible solution of the problem to be
optimized. All the paths of the whole ant colony form the solution space of the
optimization problem. The shorter path ants release more pheromones, and as time goes
on, the pheromone concentration on the shorter path increases gradually, and the
number of ants in the path is increasing. Finally, the whole ant will concentrate on the
best path under the positive feedback. At this time, the corresponding solution is the
optimal solution of the problem to be optimized.
Researchers based on the study of ACO use traveling salesman problem as a test
of performance, not others. In this paper, the basic ACO is modified accordingly, using
cloudsim platform as simulation experiment, and design, modify the corresponding
parameters, in order to design a load balancing algorithm.

1 Mathematic Model of Task scheduled

ACO is a new bionic intelligent algorithm. The first ant guides and influences the
selection of the following ants by means of pheromone, so as to achieve the goal of
cooperation among ants. The application of ant colony algorithm in the task scheduling
of cloud computing is studied in this paper.
In a data center of cloud computing, there are n virtual hosts, they are set as
,Each virtual host has a different MIPS (millions of in-
structions per second). At the same time, there are m tasks that need to be scheduled to
work on the n virtual hosts. The total amount of time spent in these tasks is:

Where indicates that the i virtual host completes the time allocated to all of its
tasks, calculated by formula (2) and formula (3).

In the above two formulas, suppose that the performance of platform I virtual host
is MIPS, and there are k tasks allocated on it, the number of instructions corresponding
to each task is (Million Instructions), so the time for each task is .
2 Ant Colony Optimization

2.1 Theory of Ant Colony Optimization

After finding the food, the ants will randomly explore the way from home to food,
and leave on the road with pheromones that are inversely proportional to the length of
the journey. Subsequent ants will be based on the concentration of pheromones, select
the nearest road, so as to achieve the group to find the best path of the results. If the
shortest path is interrupted, the ants will randomly explore and converge quickly to the
new shortest path.
The basic ACO in the TSP (traveling salesman problem) on the state as follows:
at time 0, all the ants and n city will initialize m ants randomly in a city as a starting
point, each edge between the city has an initial pheromone concentration Each
ant places its city as the first element of the tabu list. Each ant then selects the next city
to move, based on the probability function of the two parameter( ) (formula (6)),
moving from the city I to the city J, and fill out the corresponding tabu list. After a cycle
of N times, all ants go through all the cities and fill out the tabu list. After that, calculate
the path length of each ant and then update corresponding pheromone
according the formula (4) Save the shortest path that ants walk through, and get the
shortest path after one traversal And empty all the tabu tables and start the
next traversal loop. Repeat the process until the number of iterations reaches the
required number of iterations( ),get the shortest path in the global situation

where is the updated pheromone concentration, is the rate of

pheromone emission, is the sum of pheromones left by all ants.

where Q is a constant is the path length of ants.

The probability function of the first k ant moving from city I to city J is:

where means In the tabu list of K, the J city has not passed yet.
In summary, fig.1 shows the flow chart using ACO to solve the TSP problem.
 Fig.1.ACO solve TSP problem

2.2 Realization of Ant Colony Optimization

In the solution of the shortest time, m ants were placed in the initial position,
according to the rules of a certain probability choice every ant can arrange Vm for
next task, until all tasks completed, and then pheromone increment table will be updated
based on the total time. After all ants have finished their work, the best solution is
selected from all candidate allocation and the historical optimal solution, and it is
recorded, then the homologous pheromone table is updated. Then the pheromone is
corrected, simulating the ants releasing pheromone on the optimal path and the natural
volatilization of pheromone, leading the ant to find the global optimal solution. Fig.2
shows the flow chart of the algorithm implementation.
 Fig.2 the flow chart of the algorithm implementation

(1) calculation rules of pheromone increment table

After each walk, each ant leaves a certain amount of pheromones according to the
length of the path. The more ants there are in the same path, the more pheromones they
leave behind at the same time. The pheromone increment table records the size of the
pheromone emitted by the ants after each cycle. In this paper, the pheromone increment
with , calculation rules such as formula (7) for:

In the formula, the Q is expressed as a constant, and L is the total time after which
the current ant is assigned the task. Each ant will leave pheromones on its path, and in
the end, the global pheromone increment is the sum of the pheromone left by all the
ants.
(2) calculation rule of task transfer probability
When the ant chooses the virtual host for the next task, it is determined by the
pheromone left before it and a heuristic information. The probability formula for
assigning the task J to the virtual host I is shown in formula (6), where is
pheromone concentration, is the heuristic information. The formula of is
given by formula (8) and (9):

Where represents the total number of tasks allocated on table I Vm,

is the execution speed of the CPU for this Vm, is the length of the task to
be assigned, and is the interface bandwidth of this Vm, which is a constant in this
paper. are control parameters that control the relative importance of pheromone
and heuristic information.
(3) Amendment rules for pheromone tables
When all ants completing one iteration and getting an optimal solution,
pheromone table is needed to have a comprehensive revision according to the
pheromone emitted by ants and the phenomenon that simulates the natural evaporation
of residual pheromones, the correction rule is such as formula (10):

Where is the pheromone table after correction is pheromone

table before correction is the control parameter, indicates the degree of evaporation
of pheromones is the sum of pheromone increments left by all ants. The purpose
of pheromone correction is to assign as many pheromones as possible to the optimal
allocation mode, and better guide the selection of ants. The rule will not only store the
pheromones left by the ants, but will also evaporate them properly.

2.3 Algorithm parameter design

According to the above principle and various rules, this paper has designed and
stipulated a series of parameters. In the formula (7), the value of the constant Q will
affect the left ant pheromone on global pheromone influence, it will weaken the
influence of a better path on later ants if Q is too small otherwise it will faster
convergence which can easily fall into local optimal solution. According to the task
length and the MIPS value of Vm, repeated experiments were conducted, and finally
100 was selected. In the formula (6) and (10), are control parameter, where
respectively indicate the importance of pheromone and heuristic information,
and is the volatilization rate of pheromones. Studies have shone that when is 1,
the resulting optimum is better than the other. When value of is from 1 to 5, the
optimal solution quality will gradually become better, and when is more then 5, the
quality of the solution can be reduced, therefore, 5 is assigned to . value changes
between 0.3~0.9, the solution changed little, when its value is 0.5, the quality of
solution is the best, is suggested to value of 0.5[5]. In formula (9), MIPS is randomly
chosen in 100~300, and BW is the interface bandwidth. In this paper, the default value
of simulation software is taken. The task length of all tasks is also random, with a range
of 10000~50000.

3 Simulation and results

This paper uses cloudsim3.0 simulation platform to do simulation experiments.

By inheriting the DatacenterBroker class, we extend the DatacenterBrokerACO class
to ACO, and compile the Ant class and ACO class. We can regulate single ant activity
and ant colony activity separately.

3.1 Introduction of cloudsim

Cloudsim is a cloud computing simulation platform developed by Melbourne

University in Australia on the basis of Gridsim. CloudSim has several versions, and has
now grown to version 3 of CloudSim. CloudSim 3 uses a layered architecture.
Simulation layer provide a support for the configuration and simulation of virtual data
center environment simulation for cloud computing, including virtual machines,
memory, capacity, and bandwidth interface. This layer is used to study strategy of
virtual machine to the host distribution, and the implementation of virtual machine
scheduling function expansion core. The top layer is the user code layer, which provides
basic entities such as hosts, applications, virtual machines, user data and application
types, and agent scheduling policies. CloudSim 3 provides virtualization engines that
build and manage multiple virtualized tasks on data center nodes and flexibly switch
between time sharing and spatial sharing policies when virtualized services are
allocated. Cloudsim 3 is open source, users can expand the open source code by
expanding the interface, realizing their own scheduling strategy, and test and verify the
corresponding function based on specific environment and configuration.

3.2 Result of Simulation

Before the simulation starts, the paper specifies some parameters in the ant colony
algorithm, as shown in table 1:

Table.1. parameters of ACO

Q
100 1.0 5.0 0.5

In order to ensure the heterogeneity of all virtual hosts, Java is used to randomly specify
20 different MIPS hosts. The range of MIPS is 100~300, and the value of MIPS is
shown in table 2:

Table.2.the MIPS of 20 Vm
VmId 1 2 3 4 5
Mips 228 299 108 204 152
 VmId 6 7 8 9 10
Mips 275 202 278 109 195
VmId 11 12 13 14 15
Mips 208 294 196 143 140
VmId 16 17 18 19 20
Mips 209 292 194 212 255

At the same time, 100 tasks with different task lengths are generated randomly, and the
length is 10000~50000.
In order to determine the optimal solution, the number of ants is taken as half of the
number of virtual hosts, 1000 iterations, and the final total running time is
813.1402877697841. Task assignment is shown in table 3:

Table.3. task allocation

VMID TASK ID

0 27 73 97 50 16 15
1 72 0 56 58 68 94 42 14
2 9 84 43
3 3 74 46 54 13 17
4 11 49 86 24
5 90 88 98 93 47 8
6 7 71 62 30
7 63 59 22 37 55 95 87
8 38 70 77
9 51 89 75 10
10 52 36 91 18
11 85 44 53 92 45 12 19
12 33 61 99
13 29 39 83 35
14 32 78 4
15 28 66 1 26 20
16 6 21 64 67 40 82
 17 31 34 57 80 81
18 23 65 96 76 2 60 48
19 25 69 79 5 41

4 Conclusion

In this paper, a method of load balancing for Cloud Computing Based on ant colony
algorithm is proposed, and simulation experiments are carried out. The method
guarantees that all virtual hosts can receive task assignments as evenly as possible, and
that the total time of task completion is as short as possible. It can be seen that ACO
has the following characteristics in load balancing:
As the ants continue to spread pheromones to the environment, new information will
soon be updated to the environment, with dynamic characteristics.
Each ant makes its own choice according to the environment, and also sends out its
pheromones. Therefore, ACO has distributed features.
Ants share environmental information and influence environment at the same time,
so ACO has the characteristics of cooperative operation.
These characteristics are consistent with the super large scale, heterogeneous and
real-time characteristics of cloud computing, so ACO is important for load balancing.
Although ACO can solve the problem of load balancing, but also has the following
problems need further study:
Load balancing is a multi-dimensional and multi constrained QoS problem. In this
paper, the shortest time problem in load balancing is discussed, as well as the minimum
transfer and the economic principle.
ACO needs to initialize some parameters, and the choice of parameters has an
important influence on the performance and the final results. Therefore, in the cloud
computing environment, the setting of various parameters needs further study.
The search time of ACO is too long, and the convergence speed is slow. How to
combine ACO with neural network and particle swarm algorithm to improve the
performance and efficiency of the algorithm in order to meet the requirements of re-
sponse time and response speed of cloud computing.

References

1. Liu Peng. Cloud computing (Third Edition) [M]. Beijing: Publishing House of electronics
industry, 2015.
2. Li Fenghua. Research on load balancing scheduling algorithm of cloud computing resource
based on ant colony algorithm [D]. Yunnan: Yunnan University, 2013.
3. Zhang Ning. Research on task scheduling of cloud computing based on adaptive ant colony
algorithm [D]. Dalian: Dalian University of Technology, 2015.
4. Kong Lingjun, Zhang Xinghua, Chen Jianguo. Basic ant colony algorithm and its
improvement [J]. Journal of Beihua University, 2004, 5 (6):572-574.
5. Ye Zhiwei, Zheng Zhaobao. The parameters alpha, beta in the ant colony algorithm, with
an example of TSP problem [J]. Journal of Wuhan University, 2004 of P set, 29 (7):597-
601.
6. Wang Xiajun.Cloudsim, cloud computing simulation tools, research and application of [J].
micro computer applications, 2013, 29 (8):59-61.
7. Tian Qinghua.An Abstract of Ant Colony Optimization [J].Shijiazhuang: Journal of
Shijiazhuang Vocational Technology Institute,2004,16(6):37-38.
8. Zhang Hang,Luo Xiong.Status and Prospect of Ant Colony Optimization Algorithm [J].Hu-
nan:Information and Control,2004,33(3):318-324.