See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/357561579

A survey of the workload forecasting methods in cloud computing

Conference Paper · January 2022

CITATIONS READS

7 635

5 authors, including:

Shivam Kushwaha Jyoti Gupta

National Institute of Technology, Kurukshetra National Institute of Technology, Kurukshetra
3 PUBLICATIONS 7 CITATIONS 3 PUBLICATIONS 7 CITATIONS

SEE PROFILE SEE PROFILE

Deepika Saxena Ashutosh Kumar Singh

The University of Aizu Indian Institute of Information Technology Bhopal
83 PUBLICATIONS 1,074 CITATIONS 383 PUBLICATIONS 4,969 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Deepika Saxena on 04 January 2022.

The user has requested enhancement of the downloaded file.

 A survey of the workload forecasting methods in
cloud computing
Archana Yadav1, Shivam Kushwaha2, Jyoti Gupta3, Deepika Saxena4,

Ashutosh Kumar Singh5

National Institute of Technology, Kurukshetra, Haryana, India

Abstract

A few years ago, cloud computing had widely altered the way of computation
and storage. It is quite challenging for cloud service providers to maintain the
required Quality of Service (QoS) standard without violating a service level
(SLA) agreement. To improve the performance of the cloud, accurate
workload prediction plays a vital role. Cloud computing promises scalability,
on-demand service, and virtualization to improve the Quality of Service (QoS)
offered to end-users. In this paper, we present a comprehensive survey of
workload prediction approaches in cloud environments. It also highlights the
emerging challenges like resource wastage, excess power consumption, and
quality of service violations, etc.

Keywords: Cloud Computing · Workload Prediction · Scalability · Classifica-

tion

1. Introduction

Cloud computing continues as a model for IT service delivery at an astonishing speed

driven by a wide range of interactive and interactive features. It is widely used for
computation and storage over the internet for various domains like Marketing, Busi-
ness, Education, Banking, Entertainment, etc Instead of storing files on a portable
hard drive or local storage device, cloud-based storage makes it easier to store them in
a remote database. The popularity and usage of cloud computing are increasing day
by day, cloud-based applications make people highly dependent on it for their day-to-
day activities. It offers several advantages compared to traditional computer models
along the way including reduced costs and increased flexibility. Cloud computing is a
much-needed delivery of IT resources to the Internet at pay-as-you-go rates. Cloud is
present at remote locations rather than local servers. It serves several benefits like
Broad Network, Elasticity, Data Security, Scalable, On-Demand Service, Virtualiza-
2

tion, etc. Fig 1 illustrates the various types of characteristics of Cloud Computing. We
can process and access data and applications from servers via the Internet. It can also
help save our planet by providing a complete computing environment. Cloud compu-
ting can be described as a new computer-style where powerful and easy-to-use re-
sources are offered as an Internet service. Cloud computing will redesign information
technology (IT) processes and the IT market and is also a powerful computer infra-
structure for many applications. The purpose of using the cloud is to allow users to
benefit from all of these technologies, without the need for in-depth information about
each of them. The cloud aims to reduce costs and help users focus on their core busi-
ness instead of being blocked by IT barriers. Cloud computing creates privacy con-
cerns because a service provider can access data in the cloud at any time. It may alter
the information by mistake or intentionally or delete the information. The main moti-
vation is to reduce the error in workload prediction by extracting knowledge from
previous existing models so that we could identify current as well as predict future
workloads in the cloud environment.

Figure 1. Characteristics of Cloud Computing

This document presents an in-depth investigation of the technical predictions of the

project, their strategies used, and motivations to conduct them. The main contribu-
tions of this paper are as follows:

• We are describing what is cloud, what is cloud computing, and how it works.
• The benefits and highlighting emerging challenges of cloud computing.
• What is workload prediction including a survey of various existing workload pre-
diction models.
• Comparison among various existing workload prediction models in a tabular form.

The rest of the paper is organized as follows: Section 2 discusses related activities.
Comparison of different approaches is presented in Section 3 then we have provided
 3

emerging challenges in Section 4 and finally, we end up with a conclusion in section

5.

Workload Prediction

Forecasting takes information available in the present and uses it to predict the future.
Workload prediction is a technique that is used to improve efficiency and reduce the
operational cost of the cloud. Application management and resource management are
the two very basic requirements that motivate prediction. Fig. 2 illustrates the general
scenario of workload forecasting at the cloud data center. Servers receive millions of
requests sent by users. Applications are processed there as well all requests that come
at a certain time of speculation are compiled as historical data, which is later used to
predict future work. Historical data is collected and processed in advance for stand-
ardization. Then normal data transferred to a forecasting system to predict future
workload.

Figure 2: General Workload Forecasting Model

2. Related Work

2.1 Neural Network-based approaches:

Saxena et al. [1], have proposed a model of network load variability to predict the
average load between consecutive forecast periods. The variable load prediction mod-
el shows the predictable demand for demand in cloud data centers. This data is used
to train and evaluate data in the neural network predictive system. The AADE learn-
ing algorithm then trains the neural network. If the predicted workload is almost equal
to the actual workload then these data are sent to trained workload prediction models
otherwise it is again sent to the AADE learning algorithm.
Lu et al. [2], have proposed a model of a predictive load forecast to predict the future
value of work for cloud conservationists. The projection model for the proposed nov-
el, called RVL BPNN is based on the BP neural network algorithm and predicts a
4

much larger future load in terms of how to use the internal relationships between the
incoming workloads. The test results show that the proposed RVL BPNN model is
gaining clarity and efficiency.
Saxena et al. [3], have proposed the provision of energy-efficient resources and a
distribution framework to predict the changing needs of future applications. Predict-
ing the use of multiple resources simultaneously, A predictive model for the Online
Multi-Resource feed-forward Neural Network (OM-FNN) is proposed. The Tri-
adaptive Differential Evolution (TaDE) learning algorithm is designed for optimal use
of OM-FNN predictions. simultaneously with future applications, automated VM
deployment based on the integration of specific resource requirements and reduced
VM allocation enables PMs.
Kumar et al. [4], have developed a workload prediction model using a neural network
and a different evolution algorithm. .In the author's way, using a structure containing
n-p-q neurons.where n, p & q are the number of neurons in the sequence of input,
hidden and output.
A workload forecasting framework based on a neural network model with supervised
learning technique is proposed in [5]. To upgrade the network learning process a Bi-
Phase adaptive differential evolution (BaDE) learning algorithm is introduced to train
the model. The BaDE learning algorithm is used to train the proposed framework for
predicting the network load (WFNN) function. To train the neural predictive network,
historical data is actively provided by the cloud data keeper. The proposed activity is
measured by two real-world stocks, Saskatchewan HTTP and NASA.
Bi et al. [6], have presented an integrated forecasting system, provided by sound fil-
tering and data representation, known as Savitzky-Golay and Wavelet supported by
Stochastic Configuration Networks (SGW-SCN), to indicate the amount of workload
in the future.

2.2 Heuristic approaches:

Kumar et al. [7], have proposed a performance management system (SDWF) that
identifies the practice of predicting errors by incorporating the latest forecasts and
using it to increase the accuracy of other predictions. The model uses a healing meth-
od based on dark conditions in training neurons. This paper offers to create a way to
predict the burden of learning from past mistakes and deal with them appropriately. It
also develops one of the meta-heuristic optimization algorithm proposed by black-
hole scenarios to achieve the most accurate network tools.
Nguyen et al. [8], have established a predictive model for a series of novel timelines,
based on an advanced reading machine. Due to Uncertainty in the weights of in-
stalling over-the-counter learning equipment requires a large no. Hidden neurons to
achieve better results for this reason the model weight is increased. Winning the issue
is a new war of attrition based on opposition to extreme learning machines, choosing
the right input tools, and hidden bias. So in this model two algorithms are used for
extreme learning machines (ELM) and opposition warfare (OTWO). ELM was raised
to overcome the regression of media-based approaches in a single hidden layer of
supplying neural networks through two phases.
Zhu et al. [9], have raised the cache memory network (LSTM) encoder network. The
model has two components: the LSTM-based network encoder-decoder network.
 5

First, the data is encrypted In context, the vector then selects the code to produce the
middle prediction result of the output layer. Finally, the output layer displays the pre-
dictive values of the workload. The result was satisfactory in predicting mixed load-
ing in the computer environment.
2.3 Recurrent neural networks based approaches:

Gao et al. [10], have compared the making of predictable methods of independent
artwork. They suggest how to make predictions for a while before a predetermined
time point to allow for sufficient time for work planning based on the predicted work-
load. Continuing to improve the accuracy of the forecast, they introduced a combined
method of predicting the load, which begins to combine all the tasks into several stag-
es and trains the forecast model for each phase in sequence.
Kumar et al. [11], have analyzed and compared the accuracy of the assumptions for
different machine learning algorithms aimed at predicting the load of server logs. The
proposed speculation model for the comparison study was used using Linear Regres-
sion (LR), K- Nearest Neighbors (KNN), Support Vector Machine (SVM), ARMA,
ARIMA, and Support Vector Regression (SVR) for web applications to select the
appropriate algorithm depending on the characteristics of each load.
Singh et al. [12], have developed a dynamic predictive model using linear regulation,
ARIMA and supported vector regression for web applications. In addition to the
cloud-based information discussed in web applications with a new weather module
called technocrat load forecaster. In the cloud environment, the start of the VM takes
5-10 min, so the task of predicting work requires a solution in the first step to help
prepare the required VM for the incoming workload.

2.4 Hybrid Approaches :

Chandy et al. [13], have proposed machine learning methods to manage resource
allocation to cloud computing for large data processing systems, the simulation of the
proposed model using network simulator 2 enables better performance and resource
utilization at a higher cost, time, power, and memory usage. A random forest algo-
rithm used uninstall features that use the bootstrap reset process and upgrade the reso-
lution tree for all visual effects.
Saxena et al. [14], have proposed predictive resource management models in cloud
environments. In the conceptual framework for cloud resource management, m
users request different applications to be executed at a data center. Each application
had a specific resource requirement that is to be fulfilled by the data center. The
workload forecasting accuracy and virtual machine resource prediction should be
increased to ignore SLA violations.
Kumar et al. [15], have proposed a test of six different ways to predict the real-world
tracking of web servers and cloud created. All analyzes were performed three times as
three different tasks were used to measure the deviation of the predictions. A three-
step guessing error, means a complete error, means a complete error and a root means
a double error.
Kumar et al. [16], compared the predictive accuracy of the various machine learning
algorithms proposed to predict server load uploads. Line reversal is often the basic
 6

strategy used in statistical analysis where all the attributes included in the expecta-
tions are numeric.

3. Comparison table
Author, Algorithm Dataset Programming Advantage Disadvantage
Year language/Tool

Saxena et al., Auto-adaptive NASA & Python 3 Optimal Hardware de-

2020 learning based Saskatchewan prediction pendence
HTTP traces accuracy
Jing Bi et Integrated ma- Google Savitzky- Better fore- It’s never
al., 2018 chine learning workload da- Golay filter casting per- completely accu-
(SGW-SCN) tasets formance rate
Kumar et Self-directed HTTP web- MATLAB Provides It’s never
al., 2020 learning based logs from three R2017a more accurate completely accu-
different World weights of the rate
Wide Web ser network
Nguyen et Extreme learn- The internet Opposition- Good accu- Over-fitting
al., 2020 ing machine and traffic (in Meg- based tug of war racy problem
enhanced tug of abyte (from optimization
war optimization EU) & in Bytes (OTWO)
(from the UK))
Zhu et al., Attention-based Alibaba and ARIMA, Controls Require large
2019 LSTM encoder- dinda workload PSR+ EA- prediction amounts of
decoder network traces dataset GMDH accuracy memory band-
width
Jayakumar Self-optimized Azure and The inference The predic- Dependency
et al., generic workload LCG workloads & training were tion error is on long short
prediction frame- executed on a 16- very low term memory
work (LOADdy- core Intel Xeon (LSTM)
namics) Platinum 8153
CPU.
Hu et al., Elastic mecha- CPU work- Time series Easy scala- Data loss
2016 nism load time series approach, Kal- bility
man filter model
Singh et al., Technocrat ClarkNet MAE, MSE, Improves Workload is
2018 ARIMA and SVR and NASA RMSE, MAPE resource pro- non-linear
model visioning
oscillation
Qiu et al., Deep learning Workload EWMA model Good per- Require a
2016 datasets formance large amount of
data
Kumar et Artificial neural NASA and MATLAB Fast evolu- Hardware de-
al., 2018 network and adap- Saskatchewan tion pendence
tive differential servers’ HTTP
 7

evolution traces
Kumar et Long short term HTTP traces Python along Remembers Prone to over-
al., 2017 memory recurrent of, Calgary with Keras li- information fitting
neural network server, & Sas- brary for a long
(LSTM-RNN) katchewan time
server
Lu et al., Random varia- Publicly MATLAB High accu- Less secure
2016 ble learning rate available 7.14 racy
back-propagation google work-
neural network load traces
(RVLBPNN)
Linma et al., Query-based Three real- MySQL and Improves Require more
model world database PostgreSQL system per- powerful hard-
traces formance ware

4. Emerging Challenges
Cloud computing has placed many challenges in different aspects of data handling.
Some of these challenges are :
1) Decreased performance and overuse - Performance is an important factor when
considering cloud-based solutions. If cloud performance is unsatisfactory, it can drive
users away and reduce profits. Even a slight delay in loading an application or web
page can lead to a significant decrease in the percentage of users.
2) Adaptability- The word “adaptability” means the quality of being able to adjust to
new conditions. The prediction model should be adaptable to behavior changes.
3) Security and privacy – Data security is a major problem when switching to cloud
computing. The information held by the user or organization is sensitive and confi-
dential. Cloud security issues include identity theft, data breach, malware infection,
and many more that ultimately reduce trust among users of your apps
4) Workload fluctuation – Evaluate your workload and determine a pattern of factors
that influence workload.
5)Accurate workload prediction – Accuracy is a key factor in the forecasting of work-
loads and existing methods remain to produce 100% accurate results.
5. Conclusion:

Providing a variety of remote services and services to its customers is a key objective
of the cloud computing paradigm. Problems such as under-provision or over-
provision could be caused by the error of predicting that the past reduces the lead to
SLA violations and cloud performance, and this ultimately leads to the problem of
resource wastage. Various load forecasting schemes are provided in the literature
relating to the lack of accurate workload forecasting based on job history data and
handling issues such as Slashdot results, and workload fluctuations. This paper begins
by introducing the basic concepts and challenges to the burden forecasting systems.
After that, they passed the research on the proposed methods of forecasting the load
and described their main contribution, and used an algorithm to make predictions. A
comparison table is included in this paper to compare algorithms, data sets, etc.
 8

References:
[1] Deepika Saxena, Ashutosh Kumar Singh, et al. Auto-adaptive learning-based
workload forecasting in dynamic cloud environment.2020;
[2] Yau Lu, John panneerselvam, LucLiu, and Yan Wu, et al. RVLBPNN: A work-
load forecasting model for smart cloud computing.2016;
[3] Deepika Saxena, Ashutosh Kumar Singh, et al.A proactive auto-scaling and ener-
gy-efficient VM allocation framework using online multi-resource neural network
for cloud data center.2020;
[4] Jitendra Kumar, Ashutosh Kumar Singh, et al. Workload Prediction In Cloud Us-
ing Artificial Neural Network And Adaptive Differential Evolution.2018;
[5] Jitendra Kumar, Deepika Saxena, Ashutosh Kumar Singh, et al.BiPhase adaptive
learning-based neural network model for cloud datacenter workload forecast ing.2020;
[6] Jing Bi, Haitao Yuan, LiBo Zhang, Jia Zhang, et al. SGW-SCN: An integrated
machine learning approach for workload forecasting in geo-distributed cloud data
centers.2018;
[7] Jitendra Kumar, Ashutosh Kumar Singh, Rajkumar Buyya, et al. Self-directed
learning-based workload forecasting model for cloud resource management.2020;
[8] Thieu Nguyen, Bao Hoang, Giang Nyuyen, Binh Minh Nguyen, et al.A new work-
load prediction model using extreme learning machine and enhanced tug of war opti-
mization. 2020;
[9] Yonghua zhu, Weilin Zhang, Yihai Chen, and hanghao Gao, et al. A novel ap-
proach to workload prediction using attention-based LSTM encoder-decoder network
in cloud environment.2019;
[10] Jiechao Gao, Haoyu Wang, and Haiying Shen, et al. Machine Learning-Based
Workload Prediction in Cloud Computing.2020;
[11] Krishan Kumar, K. Gangadhara Rao, Suneetha Bulla, D Venkateswarulu, et al.
Forecasting of Cloud Computing Services Workload using Machine Learn ing.2021;
[12] Parminder Singh, Pooja Gupta, Kiran Jyoti, et al.TASM: technocrat ARIMA and
SVR model for workload prediction of web applications in the cloud.2018;
[13] Dr. Abraham Chandy, et al. Smart Resource usage prediction using cloud com-
puting for massive data processing system.2019;
[14] Deepika Saxena, Ashutosh Kumar Singh et al. Workload Forecasting and re-
search management models based on machine learning for cloud computing environ-
ment.2021;
[15] Jitendra Kumar, Ashutosh Kumar Singh, et al.Performance Assessment Of Time
Series Forecasting Models For Cloud Datacenter Network Workload Prediction.2020;
[16] Krishan Kumar, K. Gangadhara Rao, Suneetha Bulla, D Venkateswarulu, et al.
Forecasting of Cloud Computing Services Workload Using Machine Learing.2021;

View publication stats