A novel load balancing strategy based on node load comprehensive measuring under cloud computing environment

Abstract

Although the real-time migration of virtual machines brings about multi-functional management, the service performance will decline to a certain extent during the migration process, so shorter downtime and total migration time are very desirable. The traditional load balancing algorithm reflect the load information of each physical machine, but it does not take into account the actual case of virtual machine migration. Therefore, contraposing to the features of cloud computing, the destination node positioning strategy of cloud computing is proposed. It periodically acquires physical node’s CPU, memory, bandwidth usage and request connection and perform quantification with unified standard. Then, with computation of factors of virtual machine migration the comprehensive load value is determined finally. The approach measures the utilization rate and transfer cost of node computing resources, which can locate target nodes reasonably and efficiently, so it ensures the reasonable and efficient migration of virtual machines and relative load balancing under cloud computing environments.

Keywords

Load balancing cloud computing VMM comprehensive measure migration

1. Introduction

As the virtual machine load is allocated to the server, the user’s application load can be evenly allocated to the virtual machine on the client side [1]. Virtualization technology provides a virtualized running environment for task computing. It abstracts the hardware resources such as CPU, memory, network bandwidth, and separates the computing resources [2], to offer reasonable assignment according to different requirements of the task, which is also apparent to upper applications. There are three aspects of the existing load balancing strategies [3]: First, the weight problem between the load measures is not considered, and the load measurement is believed to be an equivalent; Second, the subjective weighting method is used to determine the weight of the load measurement when the load balance is studied. Although the subjective of host is expressed, the dynamic change of load measure of each machine is easy to be neglected under environment; In the third case, since the weight problem between the load measures is taken into account, it only considers its object weight and neglects the subjective weight among the load measures. The trigger strategy of virtual machine migration is based on the threshold of CPU. If multiple virtual machines choose the same physical node as the destination node of migration, it will result in clustering effect and the migration of virtual machine will lose its significance.

Based on above factors, this article adopts comprehensive load to measure the load status and migration price of the nodes, and propose a load balancing strategy based on comprehensive load cost. It discusses the design of each module for managing nodes including CPU, memory, bandwidth, acquisition of request connection number, formal definition, location strategy of destination node, communication mode between load management node and backup load management node, etc. Then, the procedure of the strategy is provided in detail. The approach not only considers the resource utilization of each physical node, but also reflect the dynamic load change by the status of request connection number, which describe the load status more accurately. The choice of destination nodes is determined by the load and topological distance of the node. Under the premise of guaranteeing the quality of service, it ensures the minimum cost of migration and solves the problem of load balancing.

The rest of this paper is organized as follows: Section 2 analyses the load measure factors of a physical node and introduces corresponding computing equations. Section 3 explains the process of improved load balancing strategy based on comprehensive load measure and provides the algorithm procedures for realization. Section 4 establishes the experiment environment and the load balancing tests are performed to verifies the effectiveness and advantages of the strategy when the virtual machine is migrated.

2. Node load balance measuring

To achieve resource load balancing under the cloud computing environment, multiple load measures should be considered when selecting the nodes. Based on the research of predecessors, the load measures will be considered as fully as possible. The load measures of this paper include CPU, memory, bandwidth, load change, topology distance and hard disk. Since traditional load balancing strategy is too single and incomplete, the strategy proposed in this paper adds two load metrics: the topological distance between nodes and the utilization rate of hard disk. The topological distance between nodes will have an impact on the migration cost during the migration, which needs to be considered; at the same time, the utilization of the hard disk can not be ignored, and the location of the destination node needs to locate a node with appropriate utilization of the hard disk.

Assuming set $N=\{N_{1},N_{2},\ldots N_{n}\}$ is $n$ physical nodes in the cloud environment.

Definition 1: CPU utilization rate of node

If the number of CPU of node $N_{i}(1\leqslant i\leqslant n)$ is $m$ , and the utilization rate of each CPU is $\textit{CPU}_{j}(1\leqslant j\leqslant m)$ , the CPU utilization rate of node $N_{i}$ is

$\displaystyle U_{p}=\frac{\sum\limits_{j=1}^{m}{\textit{CPU}_{j}}}{m}$ (1)

Definition 2: Memory occupancy rate of node

$M_{q}$ is occupied memory on node $N_{i}$ and $M_{t}$ is the total memory, the occupancy rate of memory is

$\displaystyle M_{q}=\frac{M_{q}}{M_{t}}$ (2)

Definition 3: Bandwidth utilization rate of node

The bandwidth utilization rate is denoted by the ratio between occupied width and total width as

$\displaystyle B_{p}=\frac{\textit{Band}_{q}}{\textit{Band}_{t}}$ (3)

where $\textit{Band}_{q}$ and $\textit{Band}_{t}$ denote the occupied width and total width respectively.

The destination node location strategy of virtual machine migration is to find the appropriate host in the physical machine and to migrate the virtual machine from source node to the destination node. The location of the destination node should not only pay attention to real-time change of node load, but also consider the location of the destination node synthetically, to reduce the total migration cost. To describe the dynamic changes in the node load more comprehensively, the node weight, the load change rate, and the topological distance between nodes are defined as following definitions.

Definition 3: Node weight

$C(N_{i})$ denotes the number of connections on $N_{i}$ , and its weight is depicted as the ratio between each node number and total connection number:

$\displaystyle w(N_{i})=\frac{C(N_{i})}{\sum\limits_{i=1}^{n}{C(N_{i})}}$ (4)

Definition 4: Load change rate

In the interval between $T_{1}$ and $T_{2}$ , the load of node may be changed, denoted by variable $\alpha$

$\displaystyle\alpha=\frac{(C(N_{i})+L(T_{2}))-(C(N_{i})+L(T_{1}))}{T_{2}-T_{1}% }=\frac{L(T_{2})-L(T_{1})}{T_{2}-T_{1}}$ (5)

where $L(T_{1})$ and $L(T_{2})$ denote the number newly added connections at time $T_{1}$ and $T_{2}$ .

Definition 5: Topological distance of node

$\displaystyle D=\sqrt{(x_{N_{i}}-x_{r})^{2}+(y_{N_{i}}-y_{r})^{2}}$ (6)

$D$ denotes the topological distance between source host where the virtual machine is located and $N_{i}$ .

3. Balancing strategy based on comprehensive load measure

3.1 Principle idea of improved load balancing strategy

In the cloud environment, the utilization of each node resource is controlled by VMM [4]. Through the proposed load balancing strategy and based on comprehensive measure, the migration destination host is selected from the nodes Then, the heavy load node in the cloud environment is migrated to the target host by the dynamic migration technology of the virtual machine, to realize the load balancing of the cloud environment. The load collection module acquires the load of physical nodes every other cycle T, including CPU, memory, bandwidth, request connection number and so on, and sends the information to the load monitoring module. The monitoring module calculates the load situation, and returns the results to the database. The database stores all the load of each node and updates them periodically; The load scheduling module is responsible for the location of the destination node, searching the database to obtain the load information according to the migration request. The virtual machine monitor initiates the migration after receiving the IP address [5]. Compared with the traditional method of system balancing by migration of the destination node, the load measure of the node is more comprehensive in this paper. It is more accurate to calculate the load of the node based on the weight of the fusion load measurement. The framework for destination node locating is depicted in Fig. 1.

Figure 1.

Destination node positioning frame.

3.2 The entire process of strategy

To evaluate the resource usage and distance cost of the physical machine more comprehensively, according to the definitions above such as CPU utilization, memory utilization, bandwidth utilization, node weight, load change rate and topological distance, this article designs the following equation to take into account all the factors to describe the comprehensive load status of node $N_{i}$ :

$\displaystyle\textit{Load}(N_{i})=\omega(N_{i})\ast D+\alpha\ast(U_{p}+M_{p}+B% _{p})$ (7)

For the load monitoring module, on the one hand, VMM provides the operation and resource utilization of all nodes including CPU, memory, network bandwidth, load change, topology distance and hard disk, and keeps its record in the specified position; on the other hand, it is the core function of the module. In this module, a load balancing strategy based on comprehensive measure is added to the module to optimize the location of the destination node. Then it send the location information to the load scheduling module, and the load scheduling module is sent to the migration source node. The implementation process of the load balancing strategy module is to control the resource utilization of each node through the load monitoring module. The load situation of each node is calculated and the best node is selected as the destination node.

The detailed process of our strategy is:

Input: virtual machines ( $VM_{i}$ ) to be migrated Output: IP address of destination node $S_{i}$ Step 1: Initialize the database; Step 2: VMM sends migration request and transfer it to load scheduling module by control interface; Step 3: If the value of $\textit{Load}(N_{i})$ is 0, assign the requests successively by the shortest distance, until $\textit{Load}(N_{i})$ of each nodes is not 0; Step 4: If not all of the value of $\textit{Load}(N_{i})$ is 0, find the node owing the smallest $\textit{Load}(N_{i})$ and send the IP address of destination to load scheduling module; Step 5: Load scheduling module send the IP address of $N_{i}$ to VMM by control interface; Step 6: VMM starts the migration; Step 7: If it arrives at updating period, the load acquisition module re-acquires the load information of each node. The load control module compute $\textit{Load}(N_{i})$ and update the database.

4. Simulations

The cloud computing platform CloudSim is used to analyze and evaluate the cloud computing load balancing strategy based on the virtual machine migration. The main evaluation index is the load balance of CPU, memory and other resources, and the average response time of the task. The migration manager and device agent monitor the load of the server and virtual machine and decide whether to migrate the virtual machine through load forecasting. The experiment sets load threshold of CPU and memory as 60%. The load balance of system CPU and memory is measured by load standard deviation, by detecting CPU or memory utilization of every host at a time. The load balancing is to control the load in the efficient running range of the server, avoiding the high load and low load running of the server, so the load balance can be measured with the system load standard deviation. It indicates that more servers are running in high efficient state. Therefore, the average response time of the task can be used to explain the load balance of the cloud computing center. This article achieves the load balancing of cloud computing system through virtual machine migration strategy, and the degree of load balancing can show the advantages of different virtual machine migration methods.

Figure 2.

Performance comparison of CMLB compared with similar strategies.

In the tests, the time period is set as $T=$ 15 s, and the single threshold of the key parameters is 90%. The threshold of the current load ratio of the system is set to 95%. Meanwhile, the response time and response rate of user requests are recorded in the experiment. The experimental results of the load balancing strategy based on comprehensive measurement (CMLB), the multi-objective combinatorial optimization algorithm (RIAL) [6] and the weighted rotation scheduling algorithm (WRR) [7] are shown in Fig. 2. Figure 2a is a response time comparison for three algorithms: when the number of connections is within 200, the response time is about 0 and the algorithms make little difference; when the number of connections is between 200 to 1100, the response time of CMLB algorithm is significantly lower than that of the other two algorithms; the algorithms are not very different when the number of connections exceeds 1100. Figure 2b compares the average response rate of three algorithms: when the number of connections is within 400, the response rate is close to 100%, and the difference between the algorithms is little; when the number of connections is between 400 to 850, the response rate of CMLB algorithm is slightly higher than that of the other two algorithms; when the number of connections exceeds 850, the algorithms have no obvious difference.

Figure 3.

STD comparison of resource utilization.

To observe the load balancing effect of different methods, we run different number of virtual machines on cloud computing environment, with the number of virtual machines increasing from 0 to 200. In the experiment, the standard deviation of CPU, memory and bandwidth resource utilization of each method are calculated to measure the equilibrium degree of resource use. In order to illustrate the effect of load balancing algorithm, we will compare with other schemes without any load balancing measures.

By adopting CMLB strategy to locate the destination host and migrate the virtual machine, the CPU utilization ratio of the node is generally more balanced, and the difference of the CPU utilization rate of every two nodes will not exceed 10%. Using the WRRL method, the utilization ratio of each node CPU is the most unbalanced. Both RIAL and WRR have a certain increase compared to the random allocation, but the difference of the CPU utilization ratio of the two nodes is still much more than 30%, and the system does not achieve real load balance. Since the initial CPU capacity of the virtual machine is larger in the experiment, the possibility of CPU resource overload is greater than memory and bandwidth, and the load balancing effect to CPU is better than the other two kinds of resources.

5. Conclusions and future work

Traditional load balancing algorithm only uses single computing resources to evaluate the load of nodes and it has defects for cloud computing environment. Based on the consideration of above factors, we use the comprehensive load to measure the load and migration cost of nodes. When the virtual machine monitor sends out the migration request, it will find the node with the least load as the destination node, to start the migration. The experimental results show the proposed algorithm shortens the total migration time while guaranteeing the load balancing of the system, and improves the migration efficiency. The optimized memory migration based on probabilistic prediction shortens the downtime and total migration time, and reduces the transmission of memory pages. However, in the idle system, the effect of the improved algorithm is not ideal. Moreover, in cloud computing scenarios, the study of physical nodes managed by a load management node without frequent migration of virtual machines, to keep the performance of management nodes, Therefore, these two problems will be the future work of this paper.

Footnotes

Acknowledgments

The authors acknowledge the Scientific Research Project of Liaoning Provincial Education Department (Grant: WJZ2016036).

References

Awad

A.I.

El-Hefnawy

N.A.

and Abdel_Kader

H.M.

, Enhanced particle swarm optimization for task scheduling in cloud computing environments, Procedia Computer Science65 (2015), 920–929.

Gopinath

P.P.G.

and Vasudevan

S.K.

, An in-depth analysis and study of load balancing techniques in the cloud computing environment, Procedia Computer Science50 (2015), 427–432.

Devi

D.C.

and Uthariaraj

V.R.

, Load balancing in cloud computing environment using improved weighted round robin algorithm for nonpreemptive dependent tasks, Scientific World Journal3 (2016), 1–14.

Yang

Wang

and Wang

, Load balancing algorithm of task scheduling in cloud computing environment based on honey bee behavior, Journal of Computer Applications31(1) (2015), 638–645.

Jin

Liu

Zhang

et al., Load balancing and task scheduling strategy for the cloud computing environments, Journal of Computational Information Systems11(2) (2015), 769–781.

Cai

Jian

et al., An adaptive approach to better load balancing in a consumer-centric cloud environment, IEEE Transactions on Consumer Electronics62(3) (2016), 243–250.

Peng

Ming

et al., A load balancing method for massive data processing under cloud computing environment, Intelligent Automation & Soft Computing23(4) (2017), 1–7.