Prediction methods for effective resource provisioning in cloud computing: A survey

Abstract

Nowadays, most of the companies are shifting from desktop PCs application to cloud based applications deployed on clouds to provide the effective services in the heterogeneous environments. But, in order to survive in such a cloud competitive market, cloud providers must reach quality of service(QoS) for their customers, otherwise losing their cloud customers to competitors. In cloud computing, providing good QoS is a main challenging task because workloads changes over a time. In Software-as-a-Service (SaaS) model, the workload of the cloud application changes continuously based on the user requests, and insufficient resource allocation to the application leads to the QoS dropping, loss of consumers and revenue. On the other side, allocating unnecessary amount of resources to the application which can lead wastage of cost and energy to maintain the resources such as datacenters, servers, cooling technology and network bandwidth etc. This issue can be solved with prediction methods, which can predict the future workload of the cloud application in terms of needed resources and allocate those resources in advance, and releasing the resources when they are not needed. This paper focuses on importance of prediction methods for effective resource provisioning system. This paper brings out a review on the state of the resource provisioning system. Finally, future trends of the prediction model are discussed.

Keywords

Cloud computing resource provisioning enterprise workload prediction methods machine learning techniques

1. Introduction

Nowadays, with the rapid development of the Information Technology (IT) based systems, several distributed techniques such as cloud computing, grid computing, social networks, wireless networks, utility computing, service oriented architecture, expert cloud, MapReduce and peer-to-peer computing etc., providing the resource sharing and data transfer. Cloud computing is the trending technology, creating the gateway among several networks and services to provide on-demand services to cloud users based on the PAYG (pay-as-you-go) model. The main idea of the cloud computing is, delivering the services like applications, storage space and platforms etc., through a web browsers. Cloud computing provides many types of cloud services such as Software-as-a-service (SaaS), Platform-as-a-Service (PaaS), Infrastructure-as-a-Service (IaaS), Network-as-a-Service (NaaS) and Expert-as-a-Service (EaaS) etc. Additionally, it’s providing features like flexibility, scalability and elasticity in the view of cloud applications.

Table 1
Prediction types

Category	Evaluation metric	Scope	Reference
Hybrid	CPU workload	IaaS	Mostafa et al. [108]
Hybrid	CPU workload	IaaS	Dang et al. [125]
Proactive	CPU workload	IaaS	Katja and Matjaz [3]
Proactive	Response time	SaaS	Mahesh Matjaz [4]
Proactive	Response time	SaaS	Rodrigo et al. [97]
Proactive	CPU workload	IaaS	Maurer et al. [142]
Proactive	CPU workload	SaaS/IaaS	Ayyoub et al. [143]
Proactive	Response time	IaaS	Liu et al. [113]
Proactive	CPU workload	IaaS	Tiziano and Mencagli [11]
Proactive	Request rate	SaaS	Woodman et al. [12]
Proactive	Response time	SaaS	Alireza et al. [13]
Proactive	Response time	SaaS/IaaS	Hong et al. [14]
Proactive	Response time	SaaS	Chirkin et al. [15]
Reactive	Response time	SaaS	Chunhong et al. [5]
Reactive	Response time	IaaS	Hancong et al. [10]
Reactive	Response time	IaaS	Emeakaroha et al. [144]

Elasticity is the key component to dynamic provisioning in the cloud computing technology. By using this feature, resource provisioning technique can scale the resources to reach the demand of cloud applications. Currently in cloud market, many service providers like Google App Engine, Microsoft Azure, Amazon EC2 and IBM bluemix etc., provide these type of scaling resources. In the view of elasticity, resource provisioning system performs scale-in or scale-out their resources based on the application workload. In this process, if scaling time is increase then it can lead to poor QoS. Here, dynamic provisioning technique works according to the resource workload behavior patterns such as service time distribution and request arrival rate. That means extra resources can be allocated during the peak periods and can be released them during less demand periods. The main challenge of dynamic provisioning technique is, forecasting the proper amount of resources to be allocated in less time to meet good QoS. This challenge can be fulfilled by using two approaches such as reactive and proactive approaches. Reactive approaches use predefined threshold values to change the resources configuration when workload reach threshold values. On the other hand, proactive approaches estimating the future workload of the each cloud application and allotting required amount of resources to applications within a time. Table 1 represents some of the related work of others to resource allocation strategies. To reach the scalability mentioned above, the proactive approaches play an essential role to estimating the correct amount of resources. But, achieving the prediction accuracy is the most challenging task in the prediction process.

Many literature reviews addressed the prediction approaches such as Huang et al. [39], Galante and Bona [38], Manvi and Shyam [8], Weingartner et al. [40], Maryam and Mohammad [1], Chana and Chana [41], Singh and Chana [42], Aceto et al. [43], Huebscher and McCann [44] etc., survey on cloud workload prediction in different manners, but not completely addressed in the view of effective resource provisioning system. This paper presents a systematic literature review on the resource provisioning issues and challenges, and prediction approaches for cloud applications. The main contributions of this paper are as follows:

(iii) (i)

A comprehensive survey is presented on the resource provisioning issues and challenges.

(ii)

The classification of prediction requirements is presented for applications such as evaluation metrics and characteristics of the prediction models.

(iii)

Finally, different types of prediction models are described followed by future trends of the prediction methods.

The rest of the paper structured as follows. In Section 2, presents several issues and challenges of resource provisioning system, and, dimensions and characteristics of the prediction models. Overview of different prediction methods are presented in Section 3. Finally, Section 4 presents, conclusion of this paper and some future trends of the prediction models are given.

Table 2

Resource management techniques

Scheme	Reference	Evaluation metric	Reliability	QoS	Delay
Provisioning	Chaisiri et al. [45]	Request rate, price	High	High	Medium
Provisioning	Fox et al. [46]	Response rate	High	High	Medium
Provisioning	Buyya and Rajiv [47]	SLA parameters	Medium	Medium	Medium
Provisioning	Dailey et al. [48]	Request rate	High	High	Medium
Provisioning	Vijayakumar et al. [49]	Virtual machines	High	High	Medium
Provisioning	Singh et al. [50]	Cluster workload	Medium	Medium	High
Allocation	Tomita and Kuribayashi [51]	Request rate	High	High	Low
Allocation	Ishakian and Sweha [52]	Cluster workload	Medium	Medium	Low
Allocation	He et al. [53]	Virtual machines	High	High	Low
Allocation	Morikawa and Ikebe [54]	Virtual machines	High	High	Low
Allocation	Mei et al. [55]	Physical machines	Medium	Medium	High
Allocation	Chiaraviglio and Matta [56]	Cluster workload	High	High	Low
Mapping	Lu and Turner [57]	Virtual network	High	Medium	Medium
Mapping	Zhang et al. [58]	Virtual node	High	High	Low
Mapping	Villegas and Sadjadi [59]	Request rate	Medium	Medium	Medium
Mapping	Nawfal et al. [60]	VM workload	Medium	Medium	Medium
Mapping	Xabriel et al. [61]	QoS metrics	Medium	Medium	Medium
Mapping	Leivadeas et al. [62]	Request rate	High	Medium	Low
Adaptation	Jung et al. [63]	Virtual machines	High	High	Low
Adaptation	Baldine et al. [64]	Virtual network	Medium	Medium	High
Adaptation	Prasad et al. [65]	Request rate	Medium	High	Medium
Adaptation	Hiltunen et al. [66]	Power consumption	Medium	Medium	Medium
Adaptation	Duong et al. [67]	Request rate	Medium	Medium	Medium
Adaptation	Zhu and Agrawal [68]	Response rate	High	High	Low

Figure 1.

The taxonomy of resource management techniques.

2. Resource provisioning in cloud computing

This section surveys the resource provisioning method issues, challenges in cloud computing area and prediction method role in the resource provisioning system. The effective resource provisioning technique must be allocated the resources while handling the workload fluctuations. The limited network resources and computing resources have to be share among cloud users effectively. While performing this action, need to be considered the consequences such as resource provisioning, resource mapping, resource adaptation and resource allocation.

For Infrastructure-as-a-service, resource management is one of challenging task. This resource management includes several jobs such as resource provisioning, resource adaptation, resource estimation, resource mapping, resource scheduling and resource discovery. Figure 1 presents the taxonomy of resource management techniques. Table 2 represents some of the relevant works to resource management techniques. In this table, evaluation metrics are used to compare under resource management techniques those are reliability, QoS and delay. The reliability metric describes the ability of the system to perform action consistently. If technique takes much time to scales the resource then reliability is low, otherwise it is high. The intention of the QoS is to provide quality services as mentioned in Service Level Agreement (SLA). Here, QoS includes error rate, availability, latency and bandwidth. The lower error rate, lower latency, higher bandwidth and higher availability offer good QoS. Finally, delay time considered in manner of initialize time of virtual machines.

Resource provisioning is the process of allocation of cloud resources to a user. Providing the efficient provisioning policies is one of the major issue in cloud. The main task of this policy is providing cloud resources to the cloud applications by following load balancing technique with high reliability mechanism. In this manner, Chaisiri et al. [45] proposed a resource scheme by using optimal cloud resource provisioning (OCRP) framework. This framework considered two metrics such as request rate and price to develop the provisioning policies. OCRP framework mainly verifies the uncertainty values of two metrics and based on that results, resources can be allocated to the users. Fox et al. [46] proposed a dynamic and automated framework which can change their parameters to achieve the accuracy and it also deal unexpected modifications of automated framework. But, it does not consider accuracy of correlation while considering different parameters. Buyya and Rajiv [47] developed a dynamic provisioning method in Aneka cloud platform to meet customer SLAs. This method mainly used for autonomic provisioning of resources to the cloud applications. At the same time, resources can be released automatically when application has less workload. But, it is difficult task for the proposed method to consider the whole cluster workload. Dailey et al. [48] developed a dynamic resource management to handle the bottleneck situation of application. This method has the capability to handle the provisioning process for web applications while optimizing the resource utilization. Authors developed this framework in EUCALYPTUS cloud platform. However, it does not consider auto scaling process of cloud resources. Vijayakumar et al. [49] proposed a Simple Earliest Deadline First (SEDF) scheduler to share the CPU capacity among all the virtual machines. In this method, authors used all virtual machines which are run on the top of the Xen hypervisor tool. Singh et al. [50] proposed a k-means clustering algorithm to determine the workload of servers. The main aim of this algorithm is, evaluating the efficiency of provisioning policies. Mainly, resources allocation can be done in three phases such as cloud selection, data center selection and server selection. At first phase, cloud resources can be categorized based on virtual machine types, prices and services etc. To maximize the benefits for users, choose the efficient cloud service those who are providing cloud resources from multiple clouds. In second phase, datacenter selection should be in the way of optimization of power consumption. Finally, third phase which is server selection can be done. This process should consider cost optimization while maintaining minimum infrastructure with quality services. Table 3 describes, some of the relevant works to resource provisioning techniques in the view of three different phases.

Table 3
Resource provisioning algorithms in the view of three phases

Phase	Reference	Algorithm	Objective	Example
Server	Greenberg et al. [72]	Bin packing	Node cost minimization	VM, CPU, memory
Server	Wang et al. [73]	Bin packing	VM maximum utilization	VM, memory
Server	Lin et al. [74]	Bin packing	Maximize CPU utilization	CPU, bandwidth
Server	He et al. [78]	Metaheuristic	Cost optimization	DC, server
Server	Jin et al. [80]	Game theory	CPU utilization	Server
Server	Luo and Qian [83]	Control theory	CPU utilization	Server
Datacenter	Zhu et al. [77]	Metaheuristic	Green computing	DC, server
Datacenter	Hassan et al. [79]	Game theory	Maximize profits	DC, server
Datacenter	Lin et al. [82]	Control theory	Power consumption	DC, server
Datacenter	Niu and Li [86]	Machine learning	Green computing	DC, server
Datacenter	Hancong et al. [10]	Ant colony	Energy efficiency	DC, server
Cloud	Legillon et al. [75]	Metaheuristic	Cost optimization	Cloud, DC, VM
Cloud	Zuo et al. [76]	Metaheuristic	CSP’s profits	Cloud, DC
Cloud	Bossche et al. [81]	Queuing theory	Cost optimization	Cloud, VM
Cloud	Seung et al. [84]	Machine learning	CPU utilization	Cloud, VM
Cloud	Wie et al. [85]	Machine learning	Cost optimization	Cloud, VM

2.1 Challenges in resource provisioning system

In the view of effective resource provisioning system, approaches should be maintain scalability methods which can allows scale-in and scale-out operations within time. This section brings out the challenges for enhancing resource provisioning system. The crucial issues that are mostly a raised with resource provisioning technique are multi-tenancy, virtualization technique, data management, interoperability, network infrastructure and application program interfaces etc.

The following challenges are mentioned, that can be able to enhance the resource provisioning system.

(1)
Developing the cloud application in the way, the application resources should be in elasticity manner to prevent the SLA violations. Proactive scaling methods can able to handle workload fluctuations so that cloud application are able to handle massive requests from users.
(2)
Allocating the resources to the application within a time to prevent the bottleneck requests while satisfying the SLAs such a less migration time of virtual machines, scalability of resources, throughput, availability, system performance and response time etc.
(3)
Developing the models which can be support energy efficiency policies. Due to the migration of virtual machines from one cloud application to another, maintenance of datacenters, cooling system and network etc., which can effects on revenues of cloud service providers (CSPs) and green computing.
(4)
Designing an algorithm that can be able to reduce the network load by using several optimization techniques such as ant colony, particle swarm optimization, genetic algorithms, neural networks and meta-heuristic etc. Balancing workload among several cloud networks is the primary step for efficient resource provisioning system.

2.2 Prediction method role in resource provisioning

Prediction method is a form of advanced analysis that uses statistical analysis techniques and machine learning algorithms to analyze historical and current data to make forecasting about behavior, future trends and activity. Here, analysis can be defined as the process of condensing large volumes of data into information so that users can understand and use. The prediction method uses machine learning techniques to obtain the future value. The machine learning programs learns from experiences (i.e., data) with respect to some class of tasks and performance measures. In machine learning techniques, at first phase it develops the learning model to learn the behavior of application. Based on the trained data, prediction models explore the future behavior of the application. Figure 2 presents the taxonomy of prediction methods.

Figure 2.

The taxonomy of prediction methods.

The main aim of the prediction method is, to predict the future workload of the each application which are hosted in cloud area (SaaS). The forecast resources for cloud application, is the important step for efficient resource provisioning system. The accuracy rate of prediction method is the essential factor for effective prediction approach. Because, allocating extra resources to the application while application has less workload leads wastage of resources and allocating insufficient resources to the application while application has massive workload leads poor QoS and SLAs violation. Also resource scaling, prevent SLAs violations and achieve QoS etc., depends on an accuracy rate of the prediction model.

Table 4

Prediction methods in the view of different dimensions

Dimensions	References	Metrics	Scalability
Future demand	Jiang et al. [91]	Request rate	High
Future demand	Weijia et al. [92]	Request rate	Medium
Future demand	Shi et al. [93]	Request rate	High
Number of PMs	Zhang et al. [87]	Execution time	High
Number of PMs	Amiri et al. [88]	Execution time	High
Number of request	Rodrigo et al. [97]	Request rate	Medium
Number of request	Woodman et al. [12]	Request rate	High
Number of requests	Liang et al. [6]	Request rate	High
Number of VMs	Hancong et al. [10]	CPU, disk	Medium
Number of VMs	Liu et al. [89]	Initialization time	High
Performance	Mohamed and Shami [18]	Execution time	High
Power consumption	Tiziano and Mencagli [11]	Throughput, latency	High
Power consumption	Nadjaran et al. [17]	Throughput	Medium
Power consumption	Dinh et al. [19]	Throughput, latency	Medium
Resource utilization	Mostafa et al. [108]	CPU, memory, disk	High
Resource utilization	Katja and Matjaz [3]	CPU, memory	High
Resource utilization	Balaji et al. [4]	CPU, memory	Medium
Resource utilization	Alireza et al. 13s	CPU, memory, disk	High
Resource utilization	Xu et al. [94]	Network	Medium
Resource utilization	Yang et al. [95]	CPU, memory	High
Resource utilization	Jheng et al. [96]	Network	Medium
Response time	Chunhong et al. [5]	Execution time	Medium
Response time	Hong et al. [14]	Execution time	High
Response time	Chirkin et al. [15]	Execution time	Medium
SLA parameters	Akindele and Samuel [90]	Response & execution time	Medium

2.3 Prediction method dimensions

Generally, the main challenge of the prediction method is to define the future behavior of hosted cloud application according to the historical data which is collected before. Table 4 presents several dimensions of the cloud application which are considered for evaluating the results. Mostafa et al. [108], Katja and Matjaz [3] and Mahesh et al. [4] considered resource utilization as a dimension of the cloud application. Authors calculated resource utilization performance such as CPUs, memory and disk workload. Mostafa et al. [108] proposed a hybrid framework to calculate resource utilization. The main aim of this method is, maximize the CPUs utilization. Chunhong et al. [5] developed an approach while considering response time as a dimension. Authors calculated execution time of each and every job for prediction purpose. Rodrigo et al. [97] developed a model for prediction in SaaS. Here, proposed model is considered request rate from each application. Hancong et al. [10] proposed a model for prediction while consider number of virtual machines as a dimension which are located in virtual machine pool and this method calculates the CPU load and Disk load of the virtual machine pool. Tiziano and Mencagli [11] proposed a proactive method and considered power consumption as a dimension. Here, authors are evaluated throughput and latency rate of the resources. Woodman et al. [12] developed a framework while consider number of requests as dimension to predict the future request rate from the individual application. Alireza et al. [13] proposed a model by using resource utilization dimension while calculates CPUs, disk and memory usage of resources. Hong et al. [14] proposed a prediction method by using response time dimension. Nadjaran et al. [17] developed a method by using power consumption dimension of datacenters. Dinh et al. [19] proposed a method for energy efficiency based on predictive optimization by using power consumption dimension. As mentioned in Table 4, each technique elaborated in Section 3.

2.4 Prediction method characteristics

This section expresses the important characteristics of prediction methods which represent significant attributes of prediction methods. The main attributes of prediction method can be summarized as follows:

2.4.1 Proactive method

The proactive method can able to update the trained model based on the present behavior of the application pattern. The proactive method can avoid workload burst situation of the application. The prediction approach must be a proactive manner. Due to maximum time of virtual machine initialization time and migration time, cloud provider can lose their users. So, the prediction approach should be able to predict the future required resources within time in a way that the cloud provider has sufficient time to provide the correct amount of resources. In reactive methods, based on the system changes by using predefined threshold values, the resources can be allocated. So, customers suffer from delay time for the resources allocation.

2.4.2 Adaptation

Elasticity mechanism is the main feature of the cloud computing technology. This mechanism can able to scale the resource configuration dynamically. Still, to enhance the scaling process, the prediction approaches role is required. Due to continuously changes of workload, the resource provisioning must be handle the autoscaling method with effectively. So, the prediction methods helps the resource provisioning system to perform the operation like scale-in or scale-out, before occurring bottleneck situation. For this process, the prediction method should learn the behavioral model of the application.

2.4.3 Time interval granularity

The first step for developing the prediction model is to identify which resource attributes must be monitored. Next, deciding the length of time interval. The prediction model should be collect proper information at given time interval. Sometimes, the long term time intervals creates huge data sets and at that time, model cannot works effectively. On the other side, short time interval details cannot be useful for effective prediction results.

2.4.4 Historical information

As mentioned in Table 4, several dimensions to be considered to allocate the resources like network bandwidth, storage resources, software resources, resources utilization, SLAs parameters, number of virtual machines and physical machines, request rate and power consumption etc. But, the effective prediction model will consider multiple dimensions information of the application. Here, selection of proper dimension(s) is the initial step for effective prediction results. Finding the correlation among resources, can be useful for an accurate result.

2.4.5 Accuracy rate

The effective prediction approaches mainly depends on an accuracy rate of the forecasted results. The effective prediction method should give the quality results which are closer to the actual results. To achieve the high accuracy rate, the prediction method should analyze the collected information thoroughly and learn the behavior of the model. Also, model can able to change pattern of the model based on the current situation.

In summary, the characteristics of the prediction methods gives an information about important attributes of the prediction model. Satisfying the all characteristics as it is mentioned above, those prediction methods can makes the accurate results.

Table 5
Machine learning techniques used for the prediction process

Method	Strength	Weakness	Reference
AR, MA, ARMA, ARIMA	Simple, useful for time series	Continuous retraining	Rodrigo et al. [97], Zhu and Agrawal [98], Bonvin et al. [99], Yang et al. [100],Caron et al. [101], Islam et al. [102],Zhang et al. [103], Tran et al. [104], Roy et al. [105], Tang et al. [106], Fang et al. [107], Groep et al. [2]
Reinforcement learning	No need of domainknowledge	Poor scalability in case of long term sampling	Mostafa et al. [108], Kaelbling et al. [109], Rajavel and Thangarathanam [110], Bahrpeyma et al. [111], Barrett et al. [112], Liu et al. [113], Xu et al. [114], Yau et al. [115], Amiri [1]
Markov model	Consider previous state results	Scalability is low in resource adaptation	Mostafa et al. [108], Qavami et al. [116],Khan and Anerousis [117], Li and Cheng [118], Gong et al. [119],Kalantari and Akbari [120], Cortez et al. [121]
K-nearest neighbor	Training time is less	Computational cost	Frank et al. [122], Troncoso et al. [123],Imandoust and Bolandraftar [124], Ban et al. [33]
Fuzzy logic	Classifications based on rules	Time taken process for classifications	Dang et al. [125], Hluch’y [126], Dutreilh et al. [127], Hasan et al. [128], Egrioglu et al. [129], Vazquez et al. [130], Amiri et al. [88]
Bayesian model	Probabilities for predictions	Scalability is low in resource adaptation	Kirshna and Manvi [34], Kondoa and Cirne [131]
Neural networks	Multi-step prediction	Time taken for largesample data	Islam et al. [102], Sladescu et al. [132], Chen et al. [133], Garg et al. [134], Duan et al. [135]
Support vector machine	Useful for classifications	Not suitable forlarge datasets	Sapankevych and Sankar [136], Bankole and Ajila [137], Cao [138], Chen et al. [139], Donate et al. [140], Saadatfar [141]

3. Overview of prediction methods

This section presents a general classification of proposed models, frameworks and techniques of different prediction methods. According to the several surveys on prediction models, prediction models can be categorized into four groups. Such as table driven models, control theory models, queuing theory and machine learning methods. In table driven models [1], each metric values stored in a table and considered for prediction evaluation. In control theory models [7], the model can be able to control the shared resources among cloud applications. In queuing theory models [7], the resources can be allocated via queuing networks. Here, predicted values stored in queues to allocate the resources to cloud applications. Finally machine learning techniques [1], works based on the behavior of the application by using collected historical information. But, this section mainly concentrates on several machine learning techniques, different proposed models and frameworks etc. The machine learning techniques used to predict the application pattern in different manner. Here, machine learning methods not only used to forecast the future behavior of application, additionally that are able to predict in different ways such as response time of jobs, SLA violations, request rate, power consumption factors and resources utilization etc. Table 5 presents, some of the related works to machine learning techniques used for the prediction process in different proposed models. The classification of different prediction methods are as follows:

3.1 Regression and moving average models

Regression analysis technique is one of the most useful statistical methods in prediction process to understand the relationship among the variables in data. Regression analysis is used to predict a dependent variable value $Y$ from an independent variables $X$ . The basic relationship between $X$ and $Y$ is expressed by Eq. (1). Here, $\beta_{0}$ and $\beta_{1}$ referred as coefficient factors and $\varepsilon$ referred as noise term. In the case of predicting dependent variable value $Y$ from several independent variables ( $X_{1}$ , $X_{2}$ , $X_{3}$ , …) which is referred as multiple regression model.

$\displaystyle Y=\beta_{0}+\beta_{1}\times X+\varepsilon$ (1)

Moving average model is used for time series analysis by creating a sequence of means for different subsets of the data set. Like moving average model, autoregression integrated moving average (ARIMA) model that uses time series data to predict the future trend value. ARIMA model is a combination of “autoregressive” and “moving average” models. Autoregressive model examining the differences between values in the time series. This model describe the future value of variable by linear combination of past observation and random error. Equation (2), represents basic relationship between future value ( $Y_{t}$ ) and past observation. Here $c$ and $\varepsilon$ represents constant value and random error respectively.

$\displaystyle Y_{t}=c+\sum Y_{t-1}+\varepsilon$ (2)

The Regression and Moving Average models are linear predictive methods. They predict the results according to the previous time intervals by using the parameters as a dimension. In this Context, Rodrigo et al. [97] proposed a prediction model by using ARIMA model for application workload which are hosted in SaaS model. Authors are addressed the solution for dynamic provisioning policies in a way of proactive approach. Here, ARIMA model evaluates accuracy rate of future workload prediction by using a request rate of web-servers. Zhu and Agrawal et al. [98] proposed a method to forecast the memory and CPUs utilization by using autoregression moving average with exogenous inputs (ARMAX) model and control theory model. This method is used to scale the resource configuration in a vertical way, it means scaling can be done in a way of virtual resources such as virtual machine memory, virtual machine types and virtual machine cores etc., rather than allocating extra virtual machines which is referred as horizontal scaling. Bonvin et al. [99] proposed a method to scale the servers in both ways such as horizontal and vertical scaling. But, this method developed in a way of reactive approach. Similarly, Yang et al. [100] developed a framework to change the cluster resource configuration by using reactive method. Bonvin et al. [99], Yang et al. [100] developed a method by using autoregression (AR) and moving average (MA) methods. Caron et al. [101] proposed a prediction method for grid workload by using moving average model. Islam et al. [102] proposed a framework with combination of linear regression and neural network methods to predict the future workload of the applications. Zhang et al. [103] proposed a method for hybrid clouds based on the ARIMA prediction approach. Tran et al. [104] developed a framework to predict the server workloads. This method applied the ARIMA approach for prediction. Roy et al. [105] used the autoregression moving average (ARMA) model to predict the resource workload while minimizing the cost. The main aim of this prediction method is maximizing the resource utilization. Tang et al. [106] proposed a method with combination of autoregression and bin packing algorithms to predict the future workload of virtual machines. Fang et al. [107] proposed a method to predict the each cloud data center in a group by using ARIMA approach. Similarly, Groep et al. [2] method used to predict the initialization time of each task on cluster by using linear regression (LR) and autoregression.

Comments: Linear regression and moving average models are common engineering computation methods to analyze the time series data. In machine learning algorithms, linear regression model is very useful for prediction on time series data set. Regression based methods using the application parameters (linear data) as a dimension such as request arrival rate to predict the future request arrival rate of the cloud applications. The main advantage of the regression and moving average models are simple and calculation complexity is less. These methods are based on linear algorithms, so these models are cannot be able change their trained model and behavior of the applications. If data set is large, these models involves huge matrix operations such as multiplications, inversion etc. If matrix data is large, memory can get bottleneck for fast computations. The main disadvantages of these methods is, takes a lot of time to adapt to changes of the model when prediction error rate increase.

3.2 Reinforcement learning model

The Reinforcement Learning (RL) Model is a method of learning and it develops optimal policies on a given state. Generally, the reinforcement learning model can be considered as Markov Decision Model which is continuously submit the action on given environment for better results. In reinforcement learning model, agent can submit their actions based on the environmental status and agent gets rewards from environment. The main objective of the agent is to select an optimal policy to choose the best action that maximizes the reward function value and/or minimize the risk. As shown in Fig. 3, at time ‘t’ the agent selects an action $a_{t}$ on state $s_{t}$ to transits the state from $s_{t}$ to $s_{t+1}$ . The environment return the reward $r_{t}$ to agent.

Algorithm 1: Pseudo code of Q-learning
1:	Begin
2:	Select a random initial state $Q(s,a)$ with zero value
3:	Do while (the final state has not been reached)
4:	observe the state $s_{t}$
5:	select an action $a_{t}$
6:	apply action $a_{t}$ and receive the reward $r_{t}$
7:	compute $Q(s,a)$ value with Eq. (3) and update the Q-table
8:	select the $s_{t+1}$ as the current state
9:	End Do while until $s_{t}$ is terminal
10:	End

Figure 3.

The process of reinforcement learning.

In this context, Mostafa et al. [108] proposed a method with combination of autonomic computing and reinforcement learning models to deal with fluctuations of cloud workload. It’s a hybrid approach to predict the future workload of the application in an effective manner. In this method, at first phase authors applied Linear Regression (LR) model for analyzing the workload and in the next phase, authors applied reinforcement learning model for better prediction which is called as multi-ahead prediction. In this proposed method, used Q-learning algorithm for reinforcement learning process. Algorithm 1 presents pseudo code of the Q-learning. Q-learning is a model free learning algorithm which is does not require the knowledge of the environment. As shown in Eq. (3), Q function value can be updated every time an action is applied to the environment, where $\alpha$ and $\gamma$ represents learning rate and discount factor respectively. The optimal action for cloud environment in the next time interval was referred by updated Q-value table.

$\displaystyle Q(s,a)=Q(s,a)+\alpha\times[R(s,a)+\gamma\times\max(s_{t+1},a_{t+% 1})-Q(s,a)]$ (3)

The additional advantage of Mostafa et al. [108] model is, authors used Markov Decision Process (MDP) model to make the decisions for scale-in and/or scale-out operations. Authors compared their results with Linear Regression, ARMA and Dynamic Resource Provisioning Monitoring (DRPM) methods and their results proved that, hybrid approach gives better prediction results. Littman et al. [109] surveyed on reinforcement learning method and provides several advantages of reinforcement learning method. Rajavel and Thangarathanam [110] proposed a behavioral learning system for cloud trading negotiation market. This method used reinforcement learning approach to implement the proposed work in adaptive manner and using Markov decision process to make the decisions. Bahrpeyma et al. [111] proposed an algorithm by using reinforcement learning approach to predict the resource provisioning information in cloud virtualized datacenters. This method mainly concentrates on datacenters utilization. Barrett et al. [112] developed an algorithm for automating resource allocation process by applying reinforcement learning approach. This framework mainly concentrates on scalability of cloud applications. Liu et al. [113] proposed a method for resource provisioning policies to enhance the QoS in virtualized environments. Xu et al. [114] proposed an algorithm for autonomic cloud management by using unified reinforcement learning approach. Yau et al. [115] mentioned a reinforcement learning approach features in wireless networks. Maryam and Mohammad [1] presented reinforcement learning method features in cloud application workload management.

Comments: The main advantage of the reinforcement learning approach is, no need of domain knowledge and can able to change resource configuration easily by applying continuous actions. In reinforcement learning, using observations which are collected from the interaction with the environment, action can be applied to maximize the reward value. Reinforcement learning agent learns by updated Q-value table. But, in the case of long term sampling, this approach cannot be able to satisfy the scalability feature. The poor policy might effect on performance of the model and it needs manual setup of all state transitions and associated rewards.

Figure 4.

Markov decision process states with resource utilization.

3.3 Markov model

The Markov model composed with states and transitions. With continuous transitions, can able to change status of states. This model helps to make the decisions according to the current state of environment. Generally Markov model can be used in four ways based on their observations. This model is classified in to four categories such as Markov chain, hidden Markov mode, Markov decision process and partial observable Markov decision process. In this context, Mostafa et al. [108] proposed a hybrid approach by using Markov Decision Process (MDP) model. Markov decision process is presents with finite states and transitions among states. In proposed system, MDP model used for performing scaling operations by considering resource utilization. As shown in Fig. 4, in this model, defined three states: over-utilization, normal-utilization and under-utilization. These states represents resources are under-provisioned, resource provisioning fits the performance requirements and resources are over-provisioned respectively. Qavami et al. [116] proposed an algorithm for dynamic resource provisioning policies in cloud computing by using Markov model. Based on the current states, proposed model is used to predict the provisioning policies for applications. Khan and Anerousis [117] proposed an algorithm to predict the future time series of workload in cloud area. Here, Markov model used for analyzing previous time series. Li and Cheng [118] proposed an algorithm to predict future fuzzy time series based on the Hidden Markov Model (HMM). The main goal of the proposed method is to predict the next state based on the current state status. Gong and Gu [119] proposed an elastic resource scaling algorithm for cloud system with the help of Markov decision process model. Kalantari and Akbari [120] developed a framework to predict the grid performance by using Markov model. In this model, grid workload utilization was considered as a dimension for Markov model. Cortez et al. [121] developed a future time series of internet traffic by using neural network, Markov model and time series methods. The proposed method used neural networks method to predict the time series of internet traffic rate, later applied on Markov model on states to predict the future state.

Comments: The Markov model value is described from previous state. The Markov decision process apply the actions based on the current status of the environment. The advantage of Markov model is, it helps to take the continuous decisions according to the status of environment and also, simple to define and also can be able to rebuilt the model according to the recent observations. But, the next state of the process only based on the previous state and it is not consider sequence of states. So, it cannot be able to perform adaptability actions of resources effectively. Hidden Markov model also consider only current observations which are related to previous state of the system, but previous states only are not sufficient for determine the future state.

3.4 K-nearest neighbor

The K-Nearest Neighbor (KNN) is an algorithm which maintains all classification data and decides the predicted value of classification. The KNN algorithm mainly used for classifying the data according to the closest training examples. But, before making the classification, must be follow the two important decisions. Such as ‘k’ parameter value and distance metric value. Parameter k value describes number of neighbors will be selected for KNN algorithm. The proper choice of $k$ value has a good impact on the performance of KNN algorithm. A large $k$ value might be effect on significant prediction value due to random error. The most popular method is ‘Euclidean distance’ to find the distance metric value by subtracting the training data points. Equation (4) represents basic formulae for to find Euclidean distance value. Here, $x$ and $y$ represents training data points in data set.

$\displaystyle E(x,y)=\sqrt{\sum_{i=0}^{n}(x_{i}-y_{i})^{2}}$ (4)

In this context, Frank et al. [122] developed a prediction method for time series with help of neural networks and KNN methods. This method analyzing the all previous classifications to predict the future time series. At the first phase, the proposed model analyzing the previous classifications, later by using neural networks defined the future time series from collected each classification value. Troncoso et al. [123] developed an algorithm for time series prediction. This method used to predict the short-term electric energy demand. Imandoust and Bolandraftar [124] presented a prediction methodology for economic events by using KNN approach. Ban et al. [33] also proposed a prediction framework for financial time series with the help of KNN regression. In proposed model, regression method is used to predict the time series for each classification.

Comments: The main advantage of K-Nearest Neighbor is, classifications of time series and also helps to identify the predicted values of classification. For KNN method training time is very less so, it is to understand and simple to implement. It can also works with multiclass data sets. But, sometimes due to the many classifications computational cost will be increase. KNN algorithm will not consider explicit training data before the classification. Due to that reason abstract the data is made upon classification. So that, KNN is called as a lazy learning. KNN method goes through all data points for individual classification which leads to the expensive and it gives significant performance only on smaller data set which is don’t have many features.

Figure 5.

Basic structure of fuzzy logic control system.

3.5 Fuzzy logic

The fuzzy classifications helps to categorize the data and can able to create the time series for those classifications. Generally, fuzzy logic is a problem solving technique based on the multivalued logic. Fuzzy logic can be used in three ways such as fuzzy rules, fuzzy set and linguistic. The fuzzy rules has if-then rules to elaborate the operation of the controller, fuzzy set has a collection of elements and define the degree varying from 0 to 1 and finally, linguistic variables defines the values in words (ex. high, medium, low). Figure 5 represents the basic structure of fuzzy logic control systems.

In this context, Luis et al. [125] proposed an algorithm for proactive cloud scaling based on the fuzzy time series. This method also used neural network, genetic algorithm and back propagation models to create the learning algorithm. This learning algorithm collects time series which are formed by fuzzifier and giving to trained model. They also used SLA violations to forecasting the results. Hluch’y et al. [126] proposed a strategy for server management to enhance the QoS with the help of fuzzy time series. In proposed method, used neural networks to predict the time series for classifications. Dutreilh et al. [127] proposed an algorithm for datacenters resource allocation with control theory and fuzzy logic. Using fuzzy logic, each datacenter load is predicted. Hasan et al. [128] proposed a methodology for autonomic cloud resource scaling. Egrioglu et al. [129] proposed a method to handle high order multivariate fuzzy time series based on Artificial Neural Networks (ANN). This method is based on the fuzzy clustering technique. At the first stage, the time series of each cluster is classified. Later, applied ANN method to predict the final result. Vazquez et al. [130] proposed an algorithm to handle dynamic resource provisioning in cloud datacenters by using prediction of time series. Vazquez et al. [88] addressed the solution for improvement techniques of provisioning methods by using fuzzy approach. In this method, predicted principal components estimates the future provisioning policies with help of fuzzy classifications.

Comments: The main feature of the fuzzy logic is, developing classifications of time series and it’s a simple process for classifications. Fuzzy logic provides the significant results for the resource provisioning. The assumption of the fuzzy logic is restrictive and is not worked for many resource workloads. Fuzzy logic can be able to incorporate their rules for resource manager, so that prediction approach can increase accuracy rate. But, for huge classifications takes more time. The inference system, fuzzy rules and membership functions should be selected in a proper manner, otherwise it impacts on accuracy rate.

3.6 Bayesian model

The Bayesian approach is the statistical model to assign the probabilities for each event based on the previously collected information. This model is a statistical classifier to estimate the probability of a given event which is belongs to a particular classification. The Bayesian model uses probability theory to classify the data. Equation (5) represents basic formula for Bayesian model. Here, $A$ and $B$ represents hypothesis and training data, respectively. $P(A)$ , $P(B)$ , $P(A/B)$ and $P(B/A)$ defines the probability of hypothesis $A$ , probability of training data $B$ , probability of $A$ given $B$ and probability of $B$ given $A$ , respectively. A classifier takes input feature vector contains feature values. Here $B$ is the feature vector with $B_{1}$ , $B_{2}$ , $B_{3}$ , …, $B_{n}$ as a feature values.

$\displaystyle P(A/B)=\frac{P(B/A)P(A)}{P(B)}$ (5)

Kirshna and Manvi [34] proposed an algorithm to predict the virtual resource workload in cloud environment. This method analyze the short term and long term resource requirements based on the Bayesian approach. Their results proved that, the proposed approach is able to forecast the virtual resources with better accuracy rate. Di et al. [131] proposed a method to predict the Google host load based on the Bayesian model with the help of optimization techniques. In proposed method, host load has considered as a dimension to predict the future workload of each host.

Comments: The Bayesian model has the ability to adjust the probability of an event when new data is arrived. The advantage of the Bayesian model is, assigning the probabilities for new actions based on the previous results. Sometimes, scalability is low while performing resource adaptation due to the wrong assumptions of probabilities. For classification purpose with small data set, Bayesian model might be not useful in order to make proper estimations of the probabilities for each class.

Figure 6.

The general structure of neural networks.

3.7 Neural networks

The Neural Networks (NN) also known as artificial neural network (ANN), analyze the records and perform the prediction process. Next, identifies the error rate in second phase of prediction and modify the prediction process model. The neural network works like artificial human nervous system for receiving, analyzing and transmitting information. This network has a three layers, they are input layer, hidden layer and output layer. First, input layer collects all the inputs to the model. These inputs can be sent to hidden layer(s) for processing. After processing, output data is made available at the output layer. As shown in Fig. 6, data processing can be done.

In this context, Islam et al. [102] proposed an algorithm to predict the resources for cloud applications. In this method, authors used linear regression and artificial neural network approaches to avoid bottleneck workloads. At first stage, by using linear regression method proposed method predicted times series of request arrival rate. Later, neural networks used to avoid bottleneck workloads of the cloud application. Sladescu et al. [132] proposed a method for prediction process by using artificial neural networks. In proposed method, resource load has considered for prediction process. Chen et al. [133] addressed the solution for prediction by using moving average and fuzzy neural network methods. Authors addressed the hybrid approach for accurate results. This method concentrates on fluctuations of workload in resources. For predicting workload fluctuations, proposed method used moving average model. Later, prediction outputs is sent to a fuzzy neural network (FNN) to improve the accuracy of results. Garg et al. [134] proposed a method to predict the CPUs utilization by using NN approach. Duan et al. [135] addresses the solution by using hybrid approach which is a combination of neural network and Bayesian methods to predict the response time. In proposed method, predicted probabilities of response time for resources. At the next stage, time series of response time has sent to neural networks to improve the results.

Comments: The advantage of NN approach is, it can be able to change prediction process based on the previous prediction results. Neural networks does not consider restrictive assumption to form of the cloud application workload. The main advantage is, it can be able to consider correlation factor when using several resources of the application. But, for large sampling data, takes more time for analyze and the back propagation is a difficult part for neural networks. Due to the hidden layer(s), neural networks does not provide any insights in to the behavioral patterns of the time series of workloads.

3.8 Support vector machine model

The Support Vector Machine (SVM) model categorize the input data into classifications based on the collected information. In machine learning, support vector machine comes under supervised learning category so that, analyzed data can be able to use for classification and regression models. The main idea of the support vector machine is, finding a hyperplane line which can divide the data set points based on the different classes. Hyperplane is a line which can classified the set of data. The significant hyperplane can be able to give a chance for new data points for classified correctly. In this context, Sapankevych and Sankar [136] presents the importance of SVM model while predicting the time series in cloud environment. In proposed method, authors addressed the regression based solution to predict the time series of workload. Bankole and Ajila [137] proposed a cloud client prediction model for resource provisioning in web application environment. The authors used regression and moving average models for prediction purpose at different stages. The individual results are combined for final results. Cao [138] proposed a model for time series forecasting by using SVM model. The proposed method classified the resource workload for each data center. This classification helps to maximize the accuracy rate of the prediction. Chen et al. [139] method used to forecasting the time series of CPU load by using neural networks and support vector machine model. The authors addressed the hybrid approach to estimate the CPU utilizations of VMs. The results shown that, combination of NN and SVM gives better results. Donate et al. [140] method also used hybrid approach which is combination of genetic algorithm and SVM model to predict the resource workload. To optimize the network utilization, genetic algorithm is used. Saadatfar et al. [141] proposed a method to predict the job failures in grid environment. This proposed method addressed the solution for job failure management in grid computing. Authors considered SLA violations as a dimension to predict the job failures.

Comments: The support vector machine model gives significant results for smaller datasets and it is more efficient due to uses a subset of training points in the decision function. The main advantage of support vector machine model is, it can able to handle nonlinear workload of the applications. The combination with optimized algorithms, can able to solve optimization problems. But, it cannot be able to provide scalability in large networks. This model cannot be support well to large datasets due to the high training time. And also, the overlapping classes might be effect on accuracy rate of prediction.

In summary, the different proposed models of the prediction methods gives an idea about importance of the prediction model while allocating the resources in cloud environment.

4. Conclusion and future trends

This paper represented a systematic literature survey on the resource provisioning methods and prediction methods in the view of cloud computing. In a similar way, surveyed on several resource provisioning issues and resource provisioning algorithms in the cloud environment. To solve the issues, prediction methods are recommended to enhance the resource provisioning management system. In this view, different machine learning techniques are surveyed for prediction process and, strength and weakness of the each prediction method are also discussed. Finally, new ideas are suggested to enhance the existing prediction approaches. The overall information presented in this paper helps the cloud researchers to understand the cloud resource provisioning management and importance of the prediction models.

The future trends of the prediction approaches are as follows:

(1)

To increase the accuracy rate of prediction, the learning model should be consider all dimensions of the cloud application rather than deciding prediction with one dimension, and every prediction method should be consider power consumption dimension as a mandatory. Otherwise, heavy power consumption of cloud infrastructure effects on green computing.

(2)

Correlation factor is the essential thing for better prediction results. In cloud environment, several resources will be there such as storage, processors, servers and networks etc., correlation among these resources, can be able to provide better results.

(3)

New prediction methods should be able to change their training models according to the workload fluctuations to reduce the error rate of prediction, and the Combination of two types of prediction categories such as reactive and proactive methods, can be able to avoid SLAs violations and maximize the CPUs utilization.

Footnotes

Authors’ Bios

K. Dinesh Kumar is currently pursuing the Ph.D., in the School of Computing Science and Engineering, VIT University, Chennai. He received his B.Tech and M.Tech degrees under JNTU-Hyderabad. He has published national and International publications to his credit and participated in various national and international conferences. His research area is cloud computing and also interests in grid computing, computer networks and machine learning.

E. Umamaheswari is an Associate Professor in School of Computing Science and Engineering, VIT University. She received ME degree and PhD in Software Engineering from Anna University, Chennai. She has published many international publications to her credit and conducted various national and international conferences. She is also a reviewer for various leading journals such as Springer, Inderscience, and IGI Global etc. She is the coordinator of Research Group ‘Software Engineering’ of VIT University Chennai campus. Her research area include software engineering, cloud computing and Internet of Things.

References

Maryam

and Mohammad

, Survey on prediction models of applications for resources provisioning in cloud, Journal of Network and Computer Applications 82 (2017), 93–113.

Groep

Templon

and Wolters

, Predicting job start times on clusters, in: Proceedings of the 2004 IEEE International Symposium on Cluster Computing and the Grid, IEEE Computer Society Press, Chicago, IL, USA, 2004, pp. 301–308.

Katja

and Matjaz

, AME-WPC: Advanced model for efficient workload prediction in the cloud, Journal of Network and Computer Applications 55 (2015), 191–201.

Balaji

Aswani

C.H.

and Subrahmanya

, Predictive Cloud resource management framework for enterprise workloads, Journal of King Saud University-Computer and Information Sciences (2016).

Chunhong

et al., An adaptive prediction approach based on workload pattern discrimination in the cloud, Journal of Network and Computer Applications 80 (2017), 35–44.

Liang

Zhang

Y.H.

and Liang

J.M.

, The placement method of resources and applications based on request prediction in cloud data center, Information Sciences 279 (2014), 735–745.

Zhang

Hejiao

and Wang

, Resource provision algorithms in cloud computing: A survey, Journal of Network and Computer Applications 64 (2016), 23–42.

Manvi

and Shyam

G.K.

, Resource management for Infrastructure as a Service (IaaS) in cloud computing: A survey, Journal of Network and Computer Applications 41 (2014), 424–440.

Sadeghi

M.A.

and Navimipour

N.J.

, Load balancing mechanisms and techniques in the cloud environments: Systematic literature review and future trends, Journal of Network and Computer Applications 71 (2016), 86–98.

10.

Duan

et al., Energy-aware scheduling of virtual machines in heterogeneous cloud computing systems, Future Generation Computer Systems 74 (2017), 142–150.

11.

Tiziano

D.M.

and Mencagli

, Proactive elasticity and energy awareness in data stream processing, Journal of Systems and Software 127 (2017), 302–319.

12.

Woodman

Hiden

and Watson

, Applications of provenance in performance prediction and data storage optimization, Future Generation Computer Systems 75 (2017), 299–309.

13.

Alireza

et al., Distribution based workload modelling of continuous queries in clouds, IEEE Transactions on Emerging Topics in Computing 5(1) (2017), 120–133.

14.

Hong

et al., Application execution time prediction for effective CPU provisioning in virtualization environment, IEEE Transactions on Parallel and Distributed Systems 28(11) (2017), 3074–3088.

15.

Chirkin

A.M.

et al., Execution time estimation for workflow scheduling, Future Generation Computer Systems 75 (2017), 376–387.

16.

Yogesh

et al., Reliability and energy efficiency in cloud computing systems: Survey and taxonomy, Journal of Network and Computer Applications 74 (2016), 66–85.

17.

Nadjaran

T.A.

et al., Renewable-aware geographical load balancing of web applications for sustainable data centers, Journal of Network and Computer Applications 83 (2017), 155–168.

18.

Mohamed

and Shami

, An evergreen cloud: Optimizing energy efficiency in heterogeneous cloud computing architectures, Vehicular Communications 9 (2017), 199–210.

19.

Dinh

et al., Energy efficiency for cloud computing system based on predictive optimization, Journal of Parallel and Distributed Computing 102 (2017), 103–114.

20.

Makaratzis

A.T.

Konstantinos

M.G.

and Tzovaras

, Energy modeling in cloud simulation frameworks, Future Generation Computer Systems 79 (2018), 715–725.

21.

Mehiar

et al., Energy-efficient resource allocation and provisioning framework for cloud data centers, IEEE Transactions on Network and Service Management 12(3) (2015), 377–391.

22.

Wei

Zhuang

and Zhang

, A three-dimensional virtual resource scheduling method for energy saving in cloud computing, Future Generation Computer Systems 69 (2017), 66–74.

23.

Jafarnejad

Rahmani

A.M.

, Ghomi and Qader

N.N.

, Load-balancing algorithms in cloud computing: A Survey, Journal of Network and Computer Applications 88 (2017), 50–71.

24.

Sadegh

A.M.

Ghobaei

and Toosi

A.N.

, Auto-scaling web applications in clouds: A cost-aware approach, Journal of Network and Computer Applications 95 (2017), 26–41.

25.

Luis

V.J.

et al., SaaS enabled admission control for MCMC simulation in cloud computing infrastructures, Computer Physics Communications 211 (2017), 88–97.

26.

Giovanni

et al., Self-managing cloud-native applications: Design, implementation, and experience, Future Generation Computer Systems 72 (2017), 165–179.

27.

Felipe

et al., A virtual machine scheduler based on CPU and I/O-bound features for energy-aware in high performance computing clouds, Computers and Electrical Engineering 56 (2016), 854–870.

28.

Elias

D.C.

et al., Dynamic auto-scaling and scheduling of deadline constrained service workloads on IaaS clouds, Journal of Systems and Software 118 (2016), 101–114.

29.

Deborah

et al., Workload modeling for resource usage analysis and simulation in cloud computing, Computers and Electrical Engineering 47 (2015), 69–81.

30.

Stelios

Bessis

and Buyya

, Self-managed virtual machine scheduling in Cloud systems, Information Sciences (2017).

31.

Changhe

M.M.

and Yang

, A survey of swarm intelligence for dynamic optimization: Algorithms and applications, Swarm and Evolutionary Computation 33 (2017), 1–17.

32.

Nabiel

A.E.

et al., Cost optimization approaches for scientific workflow scheduling in cloud and grid computing: A review, classifications, and open issues, Journal of Systems and Software 113 (2016), 1–26.

33.

Ban

Zhang

Pang

Sarrafzadeh

and Inoue

, Referential knn regression for financial time series forecasting, International Conference on Neural Information Processing, Springer, Berlin, Heidelberg, 2013.

34.

Kirshna

S.G.

and Manvi

S.S.

, Virtual resource prediction in cloud environment: A Bayesian approach, Journal of Network and Computer Applications 65 (2016), 144–154.

35.

Asghar

R.A.

Ghobaei

and Tofighy

, A learning automata-based ensemble resource usage prediction algorithm for cloud computing environment, Future Generation Computer Systems 79 (2018), 54–71.

36.

Ben

Lee

and Liang

, Reward-based Markov chain analysis adaptive global resource management for inter-cloud computing, Future Generation Computer Systems 79 (2018), 588–603.

37.

Amany

et al., Virtual machine consolidation enhancement using hybrid regression algorithms, Egyptian Informatics Journal 18(3) (2017), 161–170.

38.

Galante

and Bona

L.C.E.

, A survey on cloud computing elasticity, in: IEEE Proceedings of the Fifth International Conference on Utility and Cloud Computing, Chicago, IL, USA, 2012, pp. 263–270.

39.

Huang

and Miao

, A survey of resource management in multi-tier web applications, IEEE Communications Surveys and Tutorials 16(3), 1574–1590.

40.

Weingartner

Brascher

G.B.

and Westphall

C.B.

, Cloud resource management: A survey on forecasting and profiling models, Journal of Network and Computer Applications 47 (2015), 99–106.

41.

Singh

and Chana

, QoS-aware autonomic resource management in cloud computing: A systematic review, ACM Computing Surveys (CSUR) 48(3) (2016), 42.

42.

Singh

and Chana

, Cloud resource provisioning: Survey, status and future research directions, Knowledge and Information Systems 49(3) (2016), 1005–1069.

43.

Aceto

Botta

Donato

W.D.

and Pescape

, Cloud monitoring: A survey, Computer Networks 57(9) (2013), 2093–2115.

44.

Huebscher

M.C.

and McCann

J.A.

, A survey of autonomic computing–degrees, models, and applications, ACM Computing Surveys (CSUR) 40(3) (2008), 7.

45.

Chaisiri

Lee

and Niyato

, Optimization of resource provisioning cost in cloud computing, IEEE transactions on service computing 5(2) (2012), 67–78.

46.

Fox

et al., A view of cloud computing, in: Proceedings of the communications of the ACM 53(4) (2010), pp. 50–58.

47.

Buyya

and Rajiv

, Federated resource management in grid and cloud computing systems, Future Generation Computer System 26(5) (2011), 1189–1191.

48.

Dailey

M.N.

David

and Paul

, Adaptive resource provisioning for read intensive multi-tier applications in the cloud, Future Generation Computer System 27(3) (2011), 871–879.

49.

Vijayakumar

Zhu

and Agrawal

, Dynamic resource provisioning for data streaming applications in a cloud environment, in: Proceedings of the 2nd IEEE International Conference on Cloud Computing Technology and Science 5(6) (2010), pp. 1023–1039.

50.

Singh

Sharma

Cecchet

and Shenoy

, Autonomic mix-aware provisioning for non-stationary data center workloads, in: Proceedings of the 7th IEEE International Conference on Autonomic Computing and Communication 8(4) (2010), pp. 24–31.

51.

Tomita

and Kuribayashi

, Congestion control method with fair resource allocation for cloud computing environments, in: Proceedings of the IEEE Pacific Rim Conference on Communications, Computers and Signal Processing 10(3) (2011), pp. 67–75.

52.

Ishakian

and Sweha

, Dynamic pricing for efficient workload colocation, in: Proceedings of the 4th IEEE International Conference on utility and Cloud Computing 7(6) (2010), pp. 117–125.

53.

Guo

and Guo

, Real time elastic cloud management for limited resources, in: Proceedings of the 4th IEEE International Conference on Cloud Computing 3(6) (2011), pp. 622–629.

54.

Morikawa

and Ikebe

, Proposal and evaluation of a dynamic resource allocation method based on the load of VMs on IaaS, in: Proceedings of the 4th IFIP International Conference on New Technologies, Mobility and Security 5(6) (2011), pp. 1–6.

55.

Mei

Xing

L.L.

and Sivathanu

, Cloud computing performance measurements and analysis of network I/O applications in virtualized cloud, in: Proceedings of the 3rd IEEE International Conference on Cloud Computing, 5(6) (2010), pp. 59–66.

56.

Chiaraviglio

and Matta

, Green Co-operative green routing with energy efficient servers, in: Proceedings of 1st ACM International Conference on Energy Efficient Computing and Networking, 15(8) (2010), pp. 191–194.

57.

and Turner

, Efficient mapping of virtual networks on to a shared substrate, in: Technical Report on Washington University WUCSE-2006, 1(1) (2006), pp. 34–49.

58.

Zhang

Qian

and Wu

, An opportunistic resource sharing and topology-aware mapping framework for virtual networks, in: Proceedings of the IEEE INFOCOM, 2012, 13(4) (2012), pp. 2408–2416.

59.

Villegas

and Sadjadi

S.M.

, Mapping non-functional requirements to cloud applications, in: Proceedings of the 2011 IEEE Conference on SEKE, 8(4) (2011), pp. 527–532.

60.

Nawfal

A.M.

Ali

Hamidah

and Shamala

K.S.

, Impatient task mapping in elastic cloud using genetic algorithm, Journal of Computer Science 7(6) (2011), 877.

61.

Xabriel

Ejarque

C.J.

Sadjadi

S.M.

and Badia

R.M.

, Cloud application resource mapping and scaling based on monitoring of QoS constraints, in: Proceedings of the 2012 International Conference on Software Engineering and Knowledge Engineering 7(4) (2012), pp. 88–93.

62.

Leivadeas

Papagianni

and Papavassiliou

, Efficient resource mapping framework over networked clouds via iterated local search-based request partitioning, IEEE Transactions on Parallel and Distributed Systems 24(6) (2013), 1077–1086.

63.

Jung

Joshi

K.R.

and Hiltunen

M.A.

, Generating adaptation policies for multi-tier applications in consolidated server environments, in: Proceedings of the 2008 International Conference on Autonomic Computing, 11(4) (2008), pp. 23–32.

64.

Baldine

et al., The missing link: Putting the network in networked cloud computing, in: Proceedings of the International Conference on Virtual Computing Initiative, 8(5) (2009), pp. 212–230.

65.

Prasad

et al., Enabling performance intelligence for application adaptation in the future internet, Journal of Communications and Networks 13(6) (2011), 591–601.

66.

Hiltunen

M.A.

et al., Mistral: Dynamically managing power, performance, and adaptation cost in cloud infrastructures, in: Proceedings of the 2010 IEEE 30th International Conference on Distributed Computing Systems, 11(5) (2010), pp. 18–31.

67.

Duong

T.N.B.

and Goh

R.S.M.

, A framework for dynamic resource provisioning and adaptation in IaaS clouds, in: Proceedings of the 2011 IEEE Third International Conference on Cloud Computing Technology and Science, 5(4) (2009), pp. 312–319.

68.

Zhu

and Agrawal

, Resource provisioning with budget constraints for adaptive applications in cloud environments, in: Proceedings of the HPDC 2010, 8(3) (2010), pp. 304–307.

69.

Teng

and Magoules

, A new game theoretical resource allocation algorithm for cloud computing, in: Proceedings of the 1st International Conference on Advances in Grid and Pervasive Computing, 6(4) (2010), pp. 321–330.

70.

Sotomayor

Montero

R.S.

Llorente

I.M.

and Foster

, An open source solution for virtual infrastructure management in private and hybrid clouds, in: Proceedings of the IEEE International Conference on Internet Computing, 10(6) (2009), pp. 78–89.

71.

Hill

and Varaiya

, An algorithm for optimal service provisioning using resource pricing, in: Proceedings of the 13th IEEE International Conference on Networking for Global Communications, 1(2) (2009), pp. 368–373.

72.

Greenberg

Hamilton

Maltz

D.A.

and Patel

, The cost of a cloud: Research problems in data center networks, ACM SIGCOMM Computer Communication Review 39(1) (2008), 68–73.

73.

Wang

Niu

and Liang

, Dynamic cloud resource reservation via cloud brokerage, in: IEEE 33rd International Conference on Distributed Computing Systems (ICDCS), 2013, pp. 400–409.

74.

Lin

Peng

Liang

and Liu

, Novel resource allocation model and algorithms for cloud computing, in: 2013 Fourth International Conference on Emerging Intelligent Data and Web Technologies (EIDWT), IEEE, 2013, pp. 77–82.

75.

Legillon

et al., Cost minimization of service deployment in a multi-cloud environment, in: IEEE Congress on Evolutionary Computation (CEC), IEEE, 2013, pp. 2580–2587.

76.

Zuo

Zhang

and Tan

, Self-adaptive learning pso-based deadline constrained task scheduling for hybrid iaas cloud, IEEE Transaction on Automation Science and Engineering 11(2) (2014), 564–573.

77.

Zhu

Zheng

Zhou

and Lyu

M.R.

, Scaling service-oriented applications into geo distributed clouds, in: IEEE 7th International Symposium on Service Oriented System Engineering (SOSE), IEEE, 2013, pp. 335–340.

78.

et al., Developing resource consolidation frameworks for moldable virtual machines in clouds, Future Generation Computer System 32 (2014), 69–81.

79.

Hassan

M.M.

Hossain

M.S.

Sarkar

A.J.

and Huh

E.N.

, Cooperative game-based distributed resource allocation in horizontal dynamic cloud federation platform, Information System Frontiers 16(4) (2014), 523–542.

80.

Jin

et al., Competitive cloud resource procurements via cloud brokerage, in: IEEE 5th International Conference on Cloud Computing Technology and Science (Cloud Computing), 2 (2013), pp. 355–62.

81.

Bossche

R.V.

Vanmechelen

and Broeckhove

, Online cost-efficient scheduling of deadline-constrained workloads on hybrid clouds, Future Generation Computer System 29(4) (2013), 973–985.

82.

Lin

Liu

Wierman

and Andrew

, Online algorithms for geographical load balancing, in: International Green Computing Conference (IGCC), 2012, pp. 1–10.

83.

Luo

and Qian

, Burstiness-aware server consolidation via queuing theory approach in a computing cloud, in: IEEE 27th International Symposium on Parallel and Distributed Processing (IPDPS), IEEE, 2013, pp. 332–341.

84.

Seung

Lam

and Woo

, Cloudflex: Seamless scaling of enterprise applications into the cloud, in: Proceedings IEEE INFOCOM, IEEE, 2011, pp. 211–215.

85.

Wei

et al., Dynamic correlative vm placement for quality-assured cloud service, in: IEEE International Conference on Communications (ICC), IEEE, 2013, pp. 2573–2577.

86.

Niu

and Li

, An efficient distributed algorithm for resource allocation in large-scale coupled systems, in: Proceedings IEEE INFOCOM, IEEE, 2013, pp. 1501–1509.

87.

Zhang

et al., Dynamic energy-aware capacity provisioning for cloud computing environments, in: Proceedings of the 9th International Conference on Autonomic computing (ICAC ’12), pp. 145–154.

88.

Amiri

Feizi

M.R.

and Mohammad

, IDS fitted Q improvement using fuzzy approach for resource provisioning in cloud, Journal of Intelligent and Fuzzy Systems 32(1) (2017), 229–240.

89.

Liu

et al., Optimizing workload category for adaptive workload prediction in service clouds, in: Proceedings of the 13th International Conference on Service-Oriented Computing (ICSOC 2015), Goa, India, Springer-Verlag Berlin Heidelberg, pp. 87–104.

90.

Akindele

A.B.

and Samuel

A.A.

, Predicting cloud resource provisioning using machine learning techniques, in: Proceedings of the 26th IEEE Canadian Conference on Electrical and Computer Engineering (CCECE), Vancouver, Canada, 2013, pp. 1–4.

91.

Jiang

et al., Cloud analytics for capacity planning and instant VM provisioning, IEEE Transactions on Network and Service Management 10(3) (2013), 312–325.

92.

Weijia

Zhen

and Haipeng

, Adaptive resource provisioning for the cloud using online bin packing, IEEE Transactions on Computers 63(11) (2014), 2647–2660.

93.

Shi

et al., Prediction-based federated management of multi-scale resources in cloud, AISS: Advances in Information Sciences and Service Sciences 4(6) (2012), 324–334.

94.

D.Y.

Yang

S.L.

and Liu

R.P.

, A mixture of HMM, GA, and Elman network for load prediction in cloud-oriented data centers, Journal of Zhejiang University SCIENCE C 14(11) (2013), 845–858.

95.

Yang

et al., A new method based on PSR and EA-GMDH for host load prediction in cloud computing system, The Journal of Supercomputing 68(3) (2014), 1402–1417.

96.

Jheng

J.J.

et al., A novel VM workload prediction using Grey Forecasting model in cloud data center, in: International Conference on Information Networking, Phuket, Thailand, 2014, pp. 40–45.

97.

Rodrigo

N.C.

et al., Workload prediction using ARIMA model and its impact on cloud applications’ QoS, IEEE Transactions on Cloud Computing 3(4) (2015), 449–458.

98.

Zhu

and Agrawal

, Resource provisioning with budget constraints for adaptive applications in cloud environments, in: Proceeding of 19th ACM International Symposium on High Performance Distributed Computing, 2010, pp. 304–307.

99.

Bonvin

Papaioannou

T.G.

and Aberer

, Autonomic SLA driven provisioning for cloud applications, in: Proceeding of 11th International Symposium on Cluster, Cloud Grid Computing, 2011, pp. 434–443.

100.

Yang

Jian

L.R.

Qiu

and Li

, An extreme automation framework for scaling cloud applications, IBM Journal of Research and Development, 55(6) (2011), 8:1–8:12.

101.

Caron

Desprez

and Muresan

, Forecasting for grid and cloud computing on-demand resources based on pattern matching, in: Proceeding of 2nd IEEE International Conference on Cloud Computing Technology and Science, 2010, pp. 456–463.

102.

Islam

Keung

Lee

and Liu

, Empirical prediction models for adaptive resource provisioning in the cloud, Future Generation Computer System 28(1) (2012), 155–162.

103.

Zhang

Jiang

Yoshihira

K.K.

Chen

and Saxena

, Intelligent workload factoring for a hybrid cloud computing model, in: Proceeding of IEEE Congress Service, 2009, pp. 701–708.

104.

Tran

V.G.

Debusschere

and Bacha

, Hourly server workload forecasting up to 168 hours ahead using seasonal ARIMA model, in: Proceeding of 13th International Conference on Industrial Technology (ICIT), 2012, pp. 1127–1131.

105.

Roy

Dubey

and Gokhale

, Efficient autoscaling in the cloud using predictive models for workload forecasting, in: Proceeding of 4th International Conference on Cloud Computing, 2011, pp. 500–507.

106.

Tang

et al., Dynamic forecast scheduling algorithm for virtual machine placement in cloud computing environment, Journal of Supercomputing, 70(3), 1279–1296.

107.

Fang

et al., RPPS: A novel resource prediction and provisioning scheme in cloud data center, in: IEEE Proceedings of the Ninth International Conference on Services Computing, Honolulu, HI, USA, 2012, pp. 609–616.

108.

Mostafa

G.A.

Jabbehdari

and Pourmina

M.A.

, An autonomic resource provisioning approach for service-based cloud applications: A hybrid approach, Future Generation Computer Systems 78 (2018), 191–210.

109.

Kaelbling

L.P.

Littman

M.L.

and Moore

A.W.

, Reinforcement learning: A survey, Journal of Artificial Intelligence Research 4 (1996), 237–285.

110.

Rajavel

and Thangarathanam

, Adaptive Probabilistic Behavioural Learning System for the effective behavioural decision in cloud trading negotiation market, Future Generation Computer Systems 58 (2016), 29–41.

111.

Bahrpeyma

Haghighi

and Zakerolhosseini

, An adaptive RL based approach for dynamic resource provisioning in Cloud virtualized data centers, Computing 97(12) (2015), 1209–1234.

112.

Barrett

Howley

and Duggan

, Applying reinforcement learning towards automating resource allocation and application scalability in the cloud, Concurrency and Computation: Practice and Experience 25(12) (2013), 1656–1674.

113.

Liu

Zhang

Zhou

Zhang

and Liu

, Aggressive resource provisioning for ensuring QoS in virtualized environments, IEEE Transaction on Cloud Computing 3(2) (2015), 119–131.

114.

C.Z.

Rao

and Bu

, URL: A unified reinforcement learning approach for autonomic cloud management, Journal of Parallel Distributed Computing 72(2) (2012), 95–105.

115.

Yau

K.L.A.

Komisarczuk

and Teal

P.D.

, Reinforcement learning for context awareness and intelligence in wireless networks: Review, new features and open issues, Journal of Network and Computer Applications 35(1) (2012), 253–267.

116.

Qavami

H.R.

Jamali

Akbari

M.K.

and Javadi

, Dynamic resource provisioning in cloud computing: A Heuristic Markovian approach, in: International Conference on Cloud Computing, Springer International Publishing, 2013, pp. 102–111.

117.

Khan

and Anerousis

, Workload characterization and prediction in the cloud: A multiple time series approach, in: IEEE Network Operations and Management Symposium, Maui, HI, 2012, pp. 1287–1294.

118.

S.T.

and Cheng

Y.C.

, A stochastic HMM-based forecasting model for fuzzy time series, IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 40(5) (2010), 1255–1266.

119.

Gong

and Gu

, Wilkes, Press: Predictive elastic resource scaling for cloud systems, in: Network and Service Management (CNSM), International Conference on IEEE, 2010, pp. 9–16.

120.

Kalantari

and Akbari

, Grid performance prediction using state-space model, Concurrency and Computation: Practice and Experience 21(9) (2009), 1109–1130.

121.

Cortez

et al., Multi-scale internet traffic forecasting using neural networks and time series methods, Expert Systems 29(2) (2012), 143–155.

122.

Frank

Davey

and Hunt

, Time series prediction and neural networks, Journal of Intelligent and Robot System 31(1–3) (2001), 91–103.

123.

Troncoso

et al., Time-series prediction: Application to the short-term electric energy demand, Current Topics in Artificial Intelligence, Springer, Berlin, Heidelberg, 2004, pp. 577–586.

124.

Imandoust

S.B.

and Bolandraftar

, Application of K-nearest neighbor (KNN) approach for predicting economic events: Theoretical background, International Journal of Engineering Research and Applications 3(5) (2013), 605–610.

125.

Dang

et al., A proactive cloud scaling model based on fuzzy time series and SLA awareness, Procedia Computer Science 108 (2017), 365–374.

126.

Hluch‘y

et al., Effective computation resilience in high performance and distributed environments, Computing and Informatics 35(6) (2017), 1386–1415.

127.

Dutreilh

Moreau

Malenfant

Rivierre

and Truck

, From data center resource allocation to control theory and back, in: IEEE 3rd International Conference on Cloud Computing, 2010, pp. 410–417.

128.

Hasan

M.Z.

Magana

Clemm

Tucker

and Gudreddi

S.L.D.

, Integrated and autonomic cloud resource scaling, in: IEEE Network Operations and Management Symposium, 2012, pp. 1327–1334.

129.

Egrioglu

et al., A new approach based on artificial neural networks for high order multivariate fuzzy time series, Expert Systems with Applications, 36(7) (2009), 10589–10594.

130.

Vazquez

Krishnan

and John

, Time series forecasting of cloud data center workloads for dynamic resource provisioning, Journal of Wireless Mobile Networks, Ubiquitous Computing, and Dependable Applications (JoWUA), 6(3) (2015), 87–110.

131.

Kondoa

and Cirne

, Google host load prediction based on Bayesian model with optimized feature combination, Journal of Parallel Distributed and Computing, 74(1) (2014), 1820–1832.

132.

Sladescu

Fekete

Lee

and Liu

, Event aware workload prediction: A study using auction events, in: Proceeding of the 13th International Conference on Web Information Systems Engineering, Springer, Berlin, Heidelberg, 7651 (2012), pp. 368–381.

133.

Chen

et al., Self-adaptive prediction of cloud resource demands using ensemble model and subtractive-fuzzy clustering based fuzzy neural network, Computational Intelligence and Neuroscience 2015 (2015), pp. 17.

134.

Garg

S.K.

et al., SLA-based virtual machine management for heterogeneous workloads in a cloud datacenter, Journal of Network and Computer Applications 45 (2014), 108–120.

135.

Duan

et al., A hybrid intelligent method for performance modeling prediction of workflow activities in grids, in: Proceedings of the 9th IEEE/ACM International Symposium on Cluster Computing and the Grid, CCGRID ’2009, Shanghai, China, 2009, pp. 339–347.

136.

Sapankevych

N.I.

and Sankar

, Time series prediction using support vector machines: A survey, IEEE Computational Intelligence Magazine 4(2) (2009), 24–38.

137.

Bankole

A.A.

and Ajila

S.A.

, Cloud client prediction models for cloud resource provisioning in a multitier web application environment, in: IEEE 7th International Symposium on Service Oriented System Engineering (SOSE), San Francisco Bay, USA, 2013, pp. 156–161.

138.

Cao

, Support vector machines experts for time series forecasting, Neuro computing 51 (2003), 321–339.

139.

Chen

Yang

Dong

and Abraham

, Time-series forecasting using flexible neural tree model, Information Sciences 174(3–4) (2005), 219–235.

140.

Donate

J.P.

et al., Time series forecasting by evolving artificial neural networks with genetic algorithms, differential evolution and estimation of distribution algorithm, Neural Computer Applications 22(1) (2013), 11–20.

141.

Saadatfar

Fadishei

and Deldari

, Predicting job failures in auver grid based on workload log analysis, New Generation Computing 30(1)(2012), 73–94.

142.

Maurer

Brandic

and Sakellariou

, Adaptive resource configuration for Cloud infrastructure management, Future Generation Computer Systems 29(2) (2013), 472–487.

143.

Al-Ayyoub

Jararweh

Daraghmeh

and Althebyan

, Multi-agent based dynamic resource provisioning and monitoring for cloud computing systems infrastructure, Cluster Computing 18(2) (2015), 919–932.

144.

Emeakaroha

V.C.

Maurer

and Dustdar

, Cloud resource provisioning and SLA enforcement via LoM2HiS framework, Concurrency and Computations: Practice and Experience 25(10) (2013), 1462–1481.

145.

Dinesh

and Umamaheswari

An authenticated, secure virtualization management system in cloud computing,Asian Journal of Pharmaceutical and Clinical Research(2017), 45–48.

Prediction methods for effective resource provisioning in cloud computing: A survey

Abstract

Keywords

1. Introduction

Table 1 Prediction types

Table 3 Resource provisioning algorithms in the view of three phases

2.4 Prediction method characteristics

2.4.1 Proactive method

2.4.2 Adaptation

2.4.3 Time interval granularity

2.4.4 Historical information

2.4.5 Accuracy rate

Table 5 Machine learning techniques used for the prediction process

3.1 Regression and moving average models

3.4 K-nearest neighbor

3.6 Bayesian model

3.8 Support vector machine model

4. Conclusion and future trends

Footnotes

Authors’ Bios

References

Table 1
Prediction types

Table 3
Resource provisioning algorithms in the view of three phases

Table 5
Machine learning techniques used for the prediction process