Document not found! Please try again

Improved Virtual Machine Migration Approaches in Cloud Environment

6 downloads 15821 Views 806KB Size Report
Improved Virtual Machine migration approaches in. Cloud Environment. Anita Choudhary. Dept. of Computer. Science and. Engineering. Malaviya National.
2016 IEEE International Conference on Cloud Computing in Emerging Markets

Improved Virtual Machine migration approaches in Cloud Environment Anita Choudhary

M. C. Govil

Girdhari Singh

Lalit K. Awasthi

E.S. Pilli

Nitin Kumar

Dept. of Computer Science and Engineering Malaviya National Institute of Technology Jaipur, India 2013rcp9551@mnit .ac.in

Dept. of Computer Science and Engineering Malaviya National Institute of Technology Jaipur, India mcgovil.cse@mnit. ac.in

Dept. of Computer Science and Engineering Malaviya National Institute of Technology Jaipur, India girdharisingh@redi ffmail.com

Dept. of Computer Science and Engineering National Institute of Technology Hamirpur, India [email protected]

Dept. of Computer Science and Engineering Malaviya National Institute of Technology Jaipur, India [email protected] .in

Dept. of Computer Science and Engineering Malaviya National Institute of Technology Jaipur, India 2014PCP5402@mn it.ac.in

only 25% to the overall cost of operating a CDCs [7]. In 2013, it was estimated that approximately 260 million Watts of electricity used up by Google data centers, that much amount of electricity can give power to more than 200,000 homes continuously [8][9]. A serious concern for the research community.

Abstract—The rapid growth in cloud computing encourages the cloud provider to create more reliable, scalable and efficient cloud environment. For providing such a scalable environment, cloud provider develops large-scale Cloud Data Centers (CDC). Such CDCs consume a large amount of electrical energy for operational purpose and cooling purpose to keep the machines operating at a right temperature, but it also emits a large amount of carbon-dioxide. Thus to decrease the energy consumption, Virtual Machines (VMs) are consolidated to fewer numbers of active servers through VM migration techniques, this leads to Service Level Agreement (SLA) violation and performance degradation. Therefore to retain the energy-performance tradeoff, we should perform an optimum number of VM migrations. There are two cases when the VM migration performed: First when a hot spot is found and Second when a cold spot is found. But in both the cases we do not directly migrate the VM instead first we check whether it is required to migrate the VM. After that, the algorithm finds an appropriate destination host on which the VM will be migrated. In this paper, we implement three approaches to select destination host: First-fit, Best-fit, and Worst-fit. Finally compare the result of a number of VM migration performed, energy consumption and SLA violations with existing one.

The virtualization technology [10] enables Cloud providers to create a multiple VMs on a single server facilitating improvement in the utilization of resources and also increasing the provider's profit. VMs can be migrated across hosts to scale CDCs to improve efficiency by playing several types of resource management strategies like load balancing, power management, VM consolidation etc. As stated above energy inefficiency of CDCs need some solution. Changing idle mode of hosts to low-power mode or shift to switch off mode according to resource demands, thus eliminating the idle power of hosts which is 70% of their peak power [11], leads to great energy saving. Interactive service applications frequently experience highly changeable workload requests which in turn generates dynamic resource demand frequently thus, not only making efficient resource management in the cloud a difficult task but also leads to performance degradation if active or dynamic consolidation of VMs is used. This dynamic environment may also lead to increased response times, timeout, failures or even SLA's Violation if the resource requirements of the workload are not satisfied timely. The SLAs agreement is an agreement between service providers and the consumer, to ensure delivery of reliable Quality of Service (QoS).

Keywords—cloud computing, VM consolidation, energy consumption, virtual machine migration, Hotspot mitigation.

I.

INTRODUCTION

The thirst of sharing computing resources optimally among user's and organizations has lead the gradual evolution of distributed, cluster and grid computing and finally now, we are in the age of Cloud Computing [1]. It allows customers to provision resources on-demand, based on pay-as-you-go basis [2]. The cloud data centers (CDCs) contain thousands of computing nodes connected through a high-speed network and consuming huge amount of electricity. To guarantee service reliability and availability [3], CDCs are over-provisioned. Thus for the majority of the time, on an average 30% of cloud servers remains idle and often make use of 10–15% of their resource capacity [4] for the fulfillment of execution demand. This underutilization of resources leads to a phenomenal increase in cost and energy consumption [5][6]. It has been estimated that by 2014 about 75% cost would contribute to infrastructure and energy costs whereas IT would contribute /16 $31.00 © 2016 IEEE 978-1-5090-4573-0/16 $31.00 © 2016 IEEE DOI 10.1109/CCEM.2016.12

The live migration [12] techniques are used for dynamic consolidation of VM to the minimum number of hosts that can fulfill their resource requirements. There is a direct relationship between host consolidation and overload avoidance for reaching high-performance gain while meeting the SLAs. A simple technique to deal with overload condition in CDC is, maintain the utilization of host machines at a low value to accommodate the future workload resource requirements and ensure meeting of SLA conditions. However, it results in poor resource utilization, needless additional energy consumption. Live VM migration can be the solution to both the problems.

17

There is always some performance degradation associated with each VM migration which results in an increase in the SLA violation and the operational cost. An efficient VM migration strategy can be a good solution to deal with both overload [13] situations and energy consumption. The live VM migration need to answer the following two important questions – (i) when to migrate? and (ii) where to migrate? without performance degradation and increase in operational cost.

In all the above work, authors used the fixed upper and lower threshold, but in our approach we use a dynamic threshold to decide whether migration is performed or not. For dynamic allocation of resources based on application demands and support of green computing by keeping an optimum number of the server in working, [16] author use the concept of “skewness” to evaluate the inequality of server utilization (CPU, memory etc.). By minimizing skewness various types of workloads can be combined in a better way and it improves the overall utilization of host resources. For hotspot, mitigation selects the VM whose elimination can decrease the skewness of the host. Another issue of allocation VM to hosts in energy efficient way addressed by [17] and proposed various methods of VM to suitable host mapping. For this, they have proposed three types of VM selection policy: ‘‘the minimization of migration policy’’, ‘‘the highest potential growth policy’’ and ‘‘the random choice policy’’. While dealing with migration author analyzed that CPU utilization threshold change continuously due to frequently changing workload so in their later work [13] authors have proposed two type of adaptive techniques: Median Absolute Deviation (MAD) and Inter Quartile Range (IQR). Using these techniques host overload detection is performed as, if the current host utilization exceeds the upper threshold, then the host is considered as overloaded. In a dynamic cloud environment, adaptive threshold concept performs far better than the static threshold concept. To predict future load the authors have also proposed some methods like Local Regression (LR) and Robust Local Regression (RLR). With these methods, the host is considered overutilized when the only predicted value is more than 100%. For VM placement [17] authors proposed Modified Best Fit Decreasing (MBFD) algorithm. It sets the VM in decreasing order of CPU utilization and then allocates the VM to the host whose power increment is minimum it is a kind of bin packing algorithm. Our approach makes use of the time-series based future forecasting technique to estimate the future load of the host which helps us to find an appropriate destination host while dealing migrations and also mitigate the problem of unnecessary VM migrations.

The objective of our work is to develop an energy efficient live VM migration approach to reducing the number of VM migrations and reduce operational cost. The approach is based on future load forecasting methods and heuristic bin packing algorithms which in turn decides when and where to migrate the VM. The rest of the paper is organized as follows: section II presents the related work, section III presents a forecasting framework, section IV presents proposed work, section V presents experimental setup and results and section VI presents the conclusion. II.

RELATED WORK

Since last few years, high energy consumption by CDCs has attracted the attention of researcher’s. With respect to that, a number of energy aware and host consolidation algorithms have been developed. Two key issues have been considered: 1 load balancing for this number of load prediction algorithms have been proposed for future load forecasting, then it applies migration technique based on forecasting results to kept the PM load below the upper threshold of utilization, 2 host consolidation it reduced the number of active host by migrating the VM from underutilized PM to another PM whose load is below upper threshold. Aforementioned large scale CDC’s consumed enormous amount of electrical power, which results in high operational cost. For this [5] author proposed a resource management system that uses the concept of virtualization to reduce the energy consumption and provide QoS. In this work, authors optimize the resource utilization and perform dynamic VM migrations to reduce energy consumption and avoid hot spot. Due to changing in workload demand application performance degrade, for this [14] authors used the virtualization technique for server consolidation and address the issue that affects the application performance. The idea proposed by the authors of a fixed threshold value limits the maximum utilization of resources. The key performance metrics of VM are observed. According to these key metrics host, required consolidation is performed for maximizing the performance. There may be a chance of SLA violation if a resource surpasses the predefined threshold in such a case the system needs to migrate a VM to another host. For the same load balancing problem [15], authors have proposed two type of load balancing algorithms push and pull, that performs necessary VM migrations. When the host load is high push approach performed the VM migrate that is it work under medium to the high load condition. On the other hand, pull approach is applied when the host load is low or underutilized then it performed VM migration that is it works under medium to low load condition. The host over and under utilization detections used the CPU utilization, which was compared with predefined specific threshold value.

Different types of workloads like HPC and web have different types of QoS requirements that make the resource provisioning decision harder. For the execution of different types of workloads in the single data center for better utilization of resources [18], authors proposed an admission control and scheduling algorithm. Authors use the adaptive admission control approach and dynamic resource provisioning for better utilization of resources while considering different types of SLA’s. Different types of penalties are used for handling SLA’s violation and auto-scaling of resources are performed for meeting SLA’s. Authors used Artificial Neural Network (ANN) to predict future load demand of interactive application (web applications). Based on this prediction, the non-interactive workload (HPC applications) is scheduled on same or another host where interactive workload VMs are already scheduled. This leads to better utilization of host resources and also minimize the SLA violation. Another work proposed by [19] authors use the exponential smoothing

18

techniques for future load prediction which helps to make a decision whether VM migration performed or not. And if migration is performed then it selects an appropriate destination host based on the future load of that node. Both VM migration and destination selection used the smoothing technique. Our approach deals with changing load conditions and performs necessary VM migration. Our proposed approach uses the first fit, best fit, and worst first fit bin packing algorithms for destination host selection. III.

The value of Ft+1 is based on most recent observation Xt with a weight value α and most recent forecast Ft with a weight of 1 - α . Exponential smoothing reduces the storage problem because it requires the most recent observations, the most recent forecast, and value of α . Expanding the above equation by replacing Ft with its components: Ft+ 1 = α X t + (1 − a ) [α X t −1 + (1 − α ) Ft − 1 ] = α X t + α (1 − α ) X t −1 + (1 − α ) 2 Ft − 1 If we replace all the Ft −1 values of its components:

FORECASTING FRAMEWORK

Ft+1 = α X t + α (1 − α ) X t −1 + α (1 − α ) 2 X t − 2 + ... +

In our work forecasting is required to decide whether we have to perform VM migration or not. VM migration is one of the most crucial issues in dynamic resource management for large data centers. It is an expensive procedure because of its overheads at source and destination hosts, also it uses a large amount of bandwidth to transfer image of VM and dirty memory pages. Although it is used because it helps to distribute load across the number of physical machines to mitigate the hot spot and also reduces the number of active hosts by consolidating the VMs of the underutilized physical host. For effective VM migration forecasting techniques are used, these techniques predict future load, which helps to avoid unnecessary VM migrations.

α (1 − α ) t −1 X 1 + (1 − α ) t F1

The weights are decreasing exponentially so that it is called exponential smoothing. In our work X 1 , X 2 ,..., X t −1 is a set that contains a CPU utilization history of a physical host of most recent data values taken at equal time periods t and F1 , F 2 , ..., F t − 1 are the predicted values of CPU utilization. There are two points of concern related to SE Smoothing. The first point, to start the forecasting system we require F1. Since the value for F1 is not known, we can use either F1 = X1 or average the first four or five values in the data set and use this as the initial forecast values. Second point, the choice of α has a considerable impact on the forecast. To select the value of α first we calculate the Mean squared error (MSE) as follows:

In our work, we used Time series based [20] forecasting techniques for prediction of CPU utilization. It is a series of periodic data points, usually consisting of repeated measurements takes a continuous period of time interval. It is used to forecast future values on the basis of earlier observed data values. The exponential smoothing technique used the time-series based forecasting technique and it is a kind of weighted moving average technique.

n

M SE =

n

¦ (most recent n data values)

B.

n

(5)

n

Double Exponential (DE) Smoothing

The double exponential smoothing technique is used for handling the trends in data sets. The equations are:

Single Exponential (SE) Smoothing

(6) S t = α X t + (1 − α )( S t − 1 + bt − 1 ) 0 ≤α ≤1 (7) bt = γ ( S t − S t −1 ) + (1 − γ ) bt − 1 0 ≤γ ≤1 where α and γ are data and trend smoothing factors, and the values of this factor are obtained via nonlinear optimization technique. St corresponds to the smoothed value for time t, bt is our best estimate of the trend at time t and Xt corresponds to the raw data sequence of observations. The output of the algorithm Ft + m , an estimate of the value of X at time t+m is: (8) Ft + m = S t + m b t m >0 There are several procedures exist to set the initial values for St and bt. S1 is set to X1 and three possible equations for calculation of b1 are: (9) b1 = X 2 − X 1

It is a Weighted Moving Average method, where the weights decline exponentially, most recent observation weighted most. Suppose we want to forecast the next value of time series Xt. Our forecast is denoted by Ft. The forecast error is X t - F t . In the SE forecasting, it obtains the forecast value from the previous time period and corrects it using the forecast error. The forecast function for the next time period is: (2) Ft+1 = Ft + a(X t - Ft ) 0