A Machine Learning based Context-aware Prediction Framework for

Edge Computing Environments

Abdullah Fawaz Aljulayfi

1,2 a

and Karim Djemame

School of Computing, University of Leeds, Leeds, U.K.

Prince Sattam Bin Abdulaziz University, K.S.A.

Keywords: Edge Computing, Self-adaptive Systems, Machine Learning, Prediction Framework, Linear Regression,

Support Vector Regression, Neural Networks, Sliding Window.

Abstract: A Context-aware Prediction Framework (CAPF) can be provided through a Self-adaptive System (SAS) re-

source manager to support the autoscaling decision in Edge Computing (EC) environments. However, EC

dynamicity and workload fluctuation represent the main challenges to design a robust prediction framework.

Machine Learning (ML) algorithms show a promising accuracy in workload forecasting problems which may

vary according to the workload pattern. Therefore, the accuracy of such algorithms needs to be evaluated and

compared in order to select the most suitable algorithm for EC workload prediction. In this paper, a thorough

comparison is conducted focusing on the most popular ML algorithms which are Linear Regression (LR),

Support Vector Regression (SVR), and Neural Networks (NN) using real EC dataset. The experimental results

show that a robust prediction framework can be supported by more than one algorithm considering the EC

contextual behavior. The results also reveal that the NN outperforms LR and SVR in most cases.

1 INTRODUCTION

The EC paradigm has emerged to support the Internet

of Things (IoT) applications by pushing the computa-

tional capabilities towards the edge of the network

(Dolui and Datta, 2017; Shi and Dustdar, 2016). Such

support requires the efficient management of edge re-

sources to fulfill the IoT applications’ requirements

such as mobility and low latency. However, the EC

resource management process is not a trivial task be-

cause of the nature of EC and the rapid increase in the

number IoT devices which is estimated to be 41.6 bil-

lion devices (Framingham, 2019).

The SASs have seen a significant level of interest

in different research areas like autonomic computing

and pervasive computing and provide self-manage-

ment properties and exhibit system properties such as

self-awareness to achieve adaptation (Kavanagh et

al., 2019). They can monitor resources, state and be-

havior. Therefore, a SAS is a promising solution to

efficiently support the resource management automa-

tion in EC as it can adjust itself according to the op-

eration environment (Arcaini et al., 2015; D’Angelo,

https://orcid.org/0000-0002-2262-2340

https://orcid.org/0000-0001-5811-5263

2018; Kavanagh et al., 2019; Kramer and Magee,

2007; Krupitzer et al., 2015; Singh and Chana, 2015;

Xu and Buyya, 2019). Such adaptation can be either

proactive whereby the SAS uses the historical data to

forecast the future system behavior or changes in the

environment (Al-Dhuraibi et al., 2018; Galante and

De Bona, 2012; Moreno-vozmediano et al., 2019), re-

active whereby the system is adjusted in real-time by

continually monitoring the system behavior and oper-

ational environment, or hybrid whereby the system

uses both reactive and proactive approaches.

In a proactive adaptation, designing a robust pre-

diction framework for forecasting EC workload and

supporting auto-scaling is challenging (Delicato et

al., 2017; Gupta et al., 2017; Kaur et al., 2017; B. Liu

et al., 2020; Toczé and Nadjm-Tehrani, 2018). An EC

environment exhibits a dynamic workload and often

has limited resources. In order to design such frame-

work, a deep understanding of EC nature (e.g. work-

load patterns and users’ behavior) and a thorough in-

vestigation of the workload prediction methods, their

characteristics and impact on their accuracy are re-

quired (Islam et al., 2012; Nikravesh et al., 2015b).

Aljulayﬁ, A. and Djemame, K.

A Machine Learning based Context-aware Prediction Framework for Edge Computing Environments.

DOI: 10.5220/0010379001430150

In Proceedings of the 11th International Conference on Cloud Computing and Services Science (CLOSER 2021), pages 143-150

ISBN: 978-989-758-510-4

143

The advantages of these activities are twofold: 1) to

support the IoT applications’ Quality of Service

(QoS) by avoiding under-provisioning (Ajila and

Bankole, 2013; Aldossary and Djemame, 2018;

Calheiros et al., 2015; Kumar and Singh, 2018;

Lorido-Botran et al., 2014; Moreno-vozmediano et

al., 2019), and 2) to efficiently utilize the EC re-

sources thus avoiding over-provisioning and improv-

ing the system’s scalability.

Altogether, a timely research challenge is the de-

sign of the CAPF for forecasting EC workload. This

paper extends our previous work on the SAS Archi-

tecture (Aljulayfi and Djemame, 2019) and focuses

on the proactive adaptation support. Further, it anal-

yses a real EC dataset from Shanghai Telecom

(Sguangwang.com, 2018) in order to identify work-

load patterns and propose the most suitable workload

prediction model. Moreover, it compares the accu-

racy of the most well-known ML algorithms: LR,

SVR, and NN which includes investigating the effect

of window size. Finally, the CAPF is designed in ac-

cordance to the investigations’ results. The main con-

tributions of this paper are summarized as follows:

(C1) An analysis of a real EC workload dataset is per-

formed in order to understand the EC workload

pattern and train the ML prediction models.

(C2) A comparison of the most well-known ML al-

gorithms’ accuracy considering the window size

effect is conducted aiming towards workload

prediction framework.

(C3) A design of CAPF through SAS to support auto-

scaling decision using the most accurate and

suitable ML prediction algorithms is presented.

The remainder of this paper is organized as fol-

lows: Section 2 discusses the related work. This is fol-

lowed by Section 3 which presents the SAS architec-

ture. Section 4 shows the research methodology. The

results and discussion are in Section 5. Section 6 il-

lustrates the CAPF. Finally, the paper’s conclusion

and future work in Section 7.

2 RELATED WORK

This section discusses the proposed proactive adapta-

tion models to support auto-scaling systems, which

can be classified into resource utilization- and work-

load-based. The resource utilization-based studies

predict resource utilization e.g. CPU utilization to

support the auto-scaling decision. A considerable

body of research adopts this method. For example, a

CPU-utilization prediction model using the Regres-

sion-Markov chain (RMC) method targeting the ap-

plications’ QoS is proposed in (Li et al., 2018). The

results show that the RMC provides better accuracy

as compared to LR due to large fluctuation and ran-

domness. Some other studies adopt ML methods.

Three prediction models for CPU-utilization,

throughput, and response time for Cloud Computing

(CC) are proposed by (Bankole and Ajila, 2013) using

a synthetic linear workload. Furthermore, LR, SVR,

and NN ML methods are used. The results show that

the SVR outperforms LR and NN in predicting both

CPU-utilization and throughput whereas the NN out-

performs other methods in predicting the response

time. This work is extended in (Ajila and Bankole,

2013) by considering the random workload pattern.

In (Islam et al., 2012) a synthetic linear workload

pattern is generated in order to develop prediction

models to support scaling decisions. Moreover, this

work compares the accuracy of LR and NN with and

without the sliding window consideration. It reports

that the sliding window has a positive impact on the

models’ accuracies. The effect of the NN on the auto-

scaling decision technique is also evaluated using a

threshold and compared with SVR (Nikravesh et al.,

2014). Additionally, an investigation is conducted to

select the best proportion of the dataset split consid-

ering e.g. 50%/50% for training/testing. This work is

extended in (Nikravesh et al., 2015a) and aims to in-

vestigate the effect of different workload patterns (i.e.

growing, periodic, and unpredicted). Besides the slid-

ing window technique is considered.

A workload prediction model using SVR and NN

for growing, periodic, and unpredicted workload pat-

terns is proposed in (Nikravesh et al., 2015b). More-

over, the influence of window size on the selected al-

gorithms is considered. The adopted hypothesis is

claiming that the prediction auto-scaling system ac-

curacy can be improved by selecting the best predic-

tion algorithms based on the workload pattern. The

research is extended in (Nikravesh et al., 2017) to in-

vestigate the risk minimization principle using the

same methods and workload patterns. In addition, the

SVR, NN Multi-layer Perceptron (MLP), and MLP

with Weight Decay (MLPWD) are considered. The

NN is also adopted in (Kumar and Singh, 2018) to

develop a workload prediction model that is able to

learn the best mutation strategy along with optimal

crossover rate. The model is evaluated using two real

datasets and compared with maximum, average, and

back propagation network methods.

In (Moreno-vozmediano et al., 2019), an auto-

scaling system using the SVR model is introduced.

Besides, a performance model based on queuing the-

ory is proposed to determine the number of resources

that must be provisioned. The SVR model is com-

pared with e.g. LR method. Further, several SVR con-

CLOSER 2021 - 11th International Conference on Cloud Computing and Services Science

144

figurations are investigated considering the kernel

type. The results reveal that the SVR using different

configurations outperform the other methods. In (B.

Liu et al., 2020) both Autoregressive Moving Aver-

age (ARMA) and Elma Neural Network (ENN) are

used where the ENN is responsible for correcting the

prediction error of ARMA and providing the final

prediction value.

Most of the presented studies focus on the CC en-

vironment and use a synthetic workload. To the best

of our knowledge, this paper is the first to propose a

CAPF for the EC environments based on real EC

workload with a support of a proactive SAS. This

framework is designed thanks to a thorough compar-

ison of the most effective ML algorithms used in the

literature with consideration of the window size effect

to improve prediction accuracy.

3 SELF-ADAPTIVE SYSTEM

ARCHITECTURE

This section briefly illustrates our SAS architecture

by zooming in to show auto-scaling components only

due to the page limit. It is shown in Figure 1 where

the full version including the research roadmap can

be found in (Aljulayfi and Djemame, 2019). The SAS

uses the MAPE-based (Monitor, Analyse, Plan, Exe-

cute) loop with a focus on the Analyse activity as it is

the paper’s scope. The use of MAPE-based allows the

system to have a full and continuous management

over the operational environment thanks to MAPE

loop. Additionally, the design of the SAS architecture

aims to have a hybrid adaptation, but this paper only

focuses on the proactive side.

Figure 1: Self-adaptive system architecture.

The data analyser (i.e. analyse activity) is respon-

sible for analysing the monitoring data which is pro-

vided by Monitor activity. In order to support hybrid

adaptation, this activity is divided into two main com-

ponents as follows: 1) Context-aware Prediction

Framework (CAPF): is responsible for predicting the

number of tasks requests in the future by consuming

the historical workload that stored in Request Repos-

itory where these requests will be scheduled as con-

tainers. This component (highlighted in grey) sup-

ports C3. Further details about its internal compo-

nents is available in Section 6 as it is designed after

conducting the paper’s investigation. 2) Resource

Utilization Analyser: is responsible for reactive adap-

tation process which is used as a back-up for the

CAPF in case of events are not predicted.

4 METHODOLOGY

This section presents the methods that are used to-

wards achieving the research objectives.

4.1 Dataset Analysis

The paper makes use of the Shanghai Telcom dataset

which simulates the EC workload (Sguangwang.com,

2018) and reported in (Guo et al., 2019; Wang, Guo,

et al., 2019; Wang, Zhao, Huang, et al., 2019; Wang,

Zhao, Xu, et al., 2019). It provides six months of mo-

bile phones records accessing the Internet via base

stations which are distributed over Shanghai city. The

dataset has 7 attributes: month, date, start time, end

time, latitude, longitude, and user ID. The analysis of

the full data set shows that it has 6,952,921 records,

9739 mobile devices, and 3042 base stations. Further,

a thorough analysis of the dataset is conducted in or-

der to understand the workload patterns and mobile

phone users’ behavior. However, this section only

presents part of the workload analysis that is related

to this paper.

A preliminary data analysis revealed the workload

of the first month (i.e. June) has the lowest percentage

of records with missing data e.g. base station location.

Therefore,

decide

select

the

second

day

from

Figure 2: Workload pattern.

A Machine Learning based Context-aware Prediction Framework for Edge Computing Environments

145

the same month as it is representative for the rest days

in the same month in the sense that the overall work-

load pattern is periodic. The workload of this day is

shown in Figure 2 per minute after removing the ex-

treme outliers which hide the data pattern. From Fig-

ure 2, it can be seen clearly that the overall pattern of

the data is a fluctuation with decreasing, increasing,

and steady (fluctuating) behavior. Therefore, to pro-

pose a robust context-aware prediction model, the day

will be divided into three categories based on the

workload. These categories as shown in Figure 2 are

1) decreasing which includes late night and early

morning (red), 2) increasing which includes morning

(green), and 3) Fluctuating which includes afternoon

to evening (orange). Further, one hour from each cat-

egory is selected (i.e. 2nd, 12th, 14th hours) which

will be used in training and testing the prediction

models. The training and testing splitting percentage

will be discussed in Section 4.4.

4.2 Machine Learning Algorithms

Three of the most popular and widely used ML algo-

rithms, LR, SVR, and NN are considered. These al-

gorithms are able to predict the future workload effi-

ciently based on historical data (Baig et al., 2020;

Islam et al., 2012; C. Liu et al., 2017; Sapankevych

and Sankar, 2009).

The LR is the simplest and most widely used su-

pervised ML algorithm for prediction (Baig et al.,

2020; James et al., 2017). In this paper, the simplest

case of LR is used because we have only one input

variable. The SVR is an efficient learning method that

implements the Support Vector Machine principle but

produces continuous variable. The advantage of using

SVR is its ability to map the time-series to a higher

dimension using kernel function (Nikravesh et al.,

2017). The NN or Artificial Neural Network (ANN)

is a supervised learning algorithm that can be used for

both regression and classification problems

(Nikravesh et al., 2015b). A type of ANN is MLP

which is a feed-forward network that is used for a

range of problems including forecasting (Nikravesh

et al., 2017; Zhang et al., 1998). This network archi-

tecture is adopted in this paper because it is the most

popular and efficient network architecture that is used

for forecasting.

4.3 Sliding Window Technique

The sliding window technique uses the last 𝑛 samples

of the data feature in order to forecast the future value

https://www.cs.waikato.ac.nz/ml/weka/

(Nikravesh et al., 2017). The use of the sliding win-

dow technique is important to perform a supervised

ML when having only one feature in the dataset aim-

ing to train the prediction algorithm (Nikravesh et al.,

2015b). In this paper, the number of requests per time

unit feature is only available. Therefore, in order to

apply ML algorithms, the sliding window technique

is used. Indeed, the window size is an important factor

which has a significant influence on the ML predic-

tion accuracy. However, selecting the appropriate

window size is challenging because we have to find

the best window size that allows the model to capture

the data pattern and application behavior (Amiri and

Mohammad-khanli, 2017). This means a small win-

dow size might not be representative while a large

window size might cause overfitting (Nikravesh et al.,

2015b). Therefore, this paper aims to investigate the

effect of window size on the ML algorithms accuracy.

4.4 Experimental Design

This section presents the design of the experiments

and the overall approach. As mentioned, three predic-

tion models will be proposed, each model targets a

part of the day. In order to do so, firstly, we must find

the best splitting percentage that allows the ML algo-

rithms to capture the data pattern and relationship.

Further, this is done for each day part (i.e. decreasing,

increasing, and fluctuating). Then, based on the best

splitting percentage, the ML algorithms’ accuracy

will be evaluated using: Mean Absolute Error (MAE),

Root Mean Squared Error (RMSE), and Mean Abso-

lute Percentage Error (MAPE). This means for each

workload pattern, the prediction models are trained

and tested with consideration of the effect of the win-

dow size on the prediction accuracy. Finally, the

CAPF will be designed based on these investigations.

Table 1: SVR and NN configurations.

Metho

Paramete

Value

SVR

C1.0

Kernel RBF

timize

SMOim

rove

Learnin

rate 0.38

No. of hidden la

ers 1

Number of hidden neurons 4

Momentu

0.2

och/trainin

time 10000

As part of the experimental design, the implemen-

tation of ML algorithms and their configurations must

be considered. In this paper, all the selected algorithms

are implemented using the well-known ML tool

WEKA 3.8

. In terms of the configurations, one pre-

dictor is used in LR simplest case. For SVR and MLP

CLOSER 2021 - 11th International Conference on Cloud Computing and Services Science

146

(i.e. NN), we used the same configuration of

(Nikravesh et al., 2015a, 2015b) as shown Table 1 be-

cause we have same scenario and workload pattern.

5 RESULTS AND DISCUSSION

This section presents and discusses the experimental

results. The discussion for workload will be separate

as each workload represents a different pattern.

5.1 Percentage Splitting

Before comparing the ML models', it is important to

specify the best training duration that allows the mod-

els to capture and learn the data pattern. This section

presents the results of the experiments considering the

proportion of the dataset to include in the train split:

80/20 (i.e. 80% training and 20% testing) and 70/30

(i.e. 70% training and 30% testing). This means each

workload (i.e. decreasing, increasing, and fluctuat-

ing) is split and evaluated using these percentages.

The overall results show that the 80/20 is the best

split percentage as it allows the selected algorithms to

capture the data pattern and provide the most accurate

results. Further, the 80/20 split outperforms the 70/30

overall evaluation metrics and different window

sizes. Therefore, the 80/20 splitting percentage results

will be considered in the following sections.

5.2 Ml Algorithms Comparison

This section compares the accuracy of the ML predic-

tion algorithms considering the testing results and ad-

dresses accordingly contribution (C2).

Data with the decreasing workload reveals that

SVR outperforms both LR and NN in overall predic-

tion accuracy metrics. This can be seen clearly from

Table 2 which shows the evaluation metrics for de-

creasing workload. Additionally, the best ML predic-

tion value is also provided by the SVR when the win-

dow size is 3 using MAE and RMSE. If LR is com-

pared with NN, the LR outperforms NN in most

cases.

For the increasing workload pattern, the predic-

tion results in Table 3 show that SVR outperforms

both LR and NN in most cases; this is similar to the

decreasing workload pattern. However, by looking

closely at the results, the best prediction values over

the evaluation metrics are provided by NN when the

window size equals 9. Although SVR has better ac-

curacy in most cases, the NN provides the best accu-

racy in the increasing data pattern, thus, NN will be

adopted for this data pattern.

Table 2: MAE, MAPE, and RMSE values (decreasing).

Size

MAE MAPE RMSE

LR SVR NN LR SVR NN LR SVR NN

2 2.56 2.5 3.04 34.43 32.6 42.63 3.1 2.95 3.78

3 2.55 2.41 3.29 34.21 30.32 45.81 3.08 2.84 4.01

4 2.56 2.52 2.52 34.3 33.02 28.44 3.09 2.98 2.93

5 2.57 2.57 4.61 34.61 34.79 62.37 3.11 3.13 5.37

6 2.55 2.49 2.47 34.16 32.46 31.35 3.07 2.94 2.86

7 2.69 2.49 2.82 37.37 32.73 39.43 3.37 2.97 3.53

8 2.53 2.5 3.25 33.65 32.61 45.18 3.03 2.95 3.98

9 2.51 2.42 3.91 33.02 30.71 35.36 2.98 2.86 4.56

Table 3: MAE, MAPE, RMSE values (increasing).

Size

MAE MAPE RMSE

LR SVR NN LR SVR NN LR SVR NN

2 4.25 4.24 5.33 16.67 16.83 22.54 4.76 4.84 6.78

3 4.25 4.23 4.53 16.75 16.84 16.35 4.8 4.84 5.21

4 4.37 4.25 5.45 17.54 16.9 22.95 5.2 4.87 6.87

5 4.25 4.25 4.28 16.91 16.92 15.99 4.87 4.87 4.7

6 4.82 4.25 5.72 18.99 16.92 24.08 5.44 4.87 7.13

7 4.25 4.24 5.81 17.02 16.89 24.49 4.93 4.86 7.28

8 4.25 4.25 6.21 16.92 16.91 25.93 4.88 4.86 7.54

9 4.25 4.24 3.87 16.87 16.9 14.87 4.85 4.87 4.57

The prediction results of the fluctuating workload

are shown in Table 4. Unlike the decreasing and in-

creasing patterns, the result reveals that NN has better

accuracy as compared to both LR and SVR in most

cases. Moreover, its accuracy is the best when the

window size is 9 using MAPE and RMSE. Therefore,

the NN with window size 9 will be adopted in the

CAPF.

Table 4: MAE, MAPE, RMSE values (fluctuating).

Size

MAE MAPE RMSE

LR SVR NN LR SVR NN LR SVR NN

2 4.97 4.84 4.26 20.93 20.41 17.26 6.23 6.11 5.13

3 4.86 4.7 4.29 20.87 19.8 16.96 6.45 5.94 5.06

4 4.74 4.65 5.52 19.92 19.54 23.3 5.97 5.89 6.84

5 4.72 4.6 4.75 19.83 19.31 19.97 5.95 5.81 5.96

6 4.65 4.59 4.57 19.51 19.13 18.97 5.86 5.73 5.61

7 4.63 4.48 4.63 19.31 18.56 19.58 5.77 5.54 5.91

8 4.63 4.66 4.06 19.36 19.41 16.88 5.79 5.8 5.06

9 4.66 4.69 4.15 19.75 19.63 15.88 5.88 5.91 4.93

5.3 Sliding Window Effect

The increase in window size does not have a signifi-

cant impact on LR and SVR algorithms over all met-

rics in both decreasing and increasing data as shown

in Figures 3 and 4, respectively. To be more specific,

from Figure 3 and 4, the changes in the accuracy of

both LR and SVR are roughly steady when the win-

dow size increase. Although the SVR accuracy is al-

most steady, it provides the best accuracy when the

window size is 3 as compared to other ML algorithms

in decreasing data. Additionally, the difference be-

tween LR and SVR are neglected in the most cases of

different window size values. In terms of the increas-

ing data, the SVR accuracy seems to be steady over

A Machine Learning based Context-aware Prediction Framework for Edge Computing Environments

147

window size as shown in Figure 4. In contrast, the in-

crease of window size causes highly fluctuating NN

accuracies over MAE, MAPE, and RMSE in both de-

creasing and increasing data. Further, in the increas-

ing data, NN provides the best accuracy when the

window size equals to 9.

Figure 3: Window size effect (decreasing).

Figure 4: Window size effect (increasing).

Figure 5: Window size effect (fluctuating).

Unlike decreasing and increasing data, increasing the

window size has a positive impact on all ML algo-

rithms overall evaluation metrics in fluctuating data.

This effect can be seen clearly in Figure 5 which

shows a decreasing trend. Further, the sliding window

technique slightly improves the accuracy of both LR

and SVR. In the case of NN, it has a significant im-

pact on its accuracy over MAE, MAPE, and RMSE.

Additionally, the best accuracy is provided by NN

when the window size is 9 using MAPE and RMSE.

5.4 Results Summary

This section highlights the main findings of the ex-

periments and their position in the context of the re-

lated work. NN outperforms LR and SVR in both in-

creasing and fluctuating workloads whereas SVR out-

performs NN and LR in decreasing workload. The

reason NN exhibits the best accuracy is its ability to

capture all noise in the data whereas SVR tries to find

a smooth curve to cover them (Nikravesh et al.,

2017). Based on this logic, SVR should outperform

NN in increasing workload. However, the increasing

workload has some form of fluctuation which reduces

the SVR accuracy.

In terms of the sliding window, the results show

that for some workload patterns increasing the window

may have a significant impact on the prediction accu-

racy. For example, increasing the window size has a

positive impact on the ML algorithms in the fluctuating

workload because the large window size allows the

models to learn the relationships between features

(Islam et al., 2012; Nikravesh et al., 2015b). In con-

trast, the increase of window size does not have an im-

pact on some ML algorithms such as LR and SVR

which means that their accuracies are almost steady

over the window size values (Nikravesh et al., 2017).

6 CONTEXT-AWARE

PREDICTION FRAMEWORK

This section proposes the CAPF fulfilling contribu-

tion (C3). It is designed according to the above exper-

iments that select the best ML prediction algorithms

with consideration of the window size. Further, it is

integrated with the SAS architecture in the CAPF

component that is shown in Figure 1 (i.e. highlighted

in grey). The framework aims to forecast the future

workload that will be submitted to the EC by the IoT

devices. This framework consists of two main com-

ponents as shown in Figure 6 (highlighted in grey)

which are: 1) Context analyser: identifies the context

CLOSER 2021 - 11th International Conference on Cloud Computing and Services Science

148

of the application, including time dependence, in or-

der to select the ML model suits its workload pattern

(i.e. decreasing, increasing, and fluctuating). 2) Algo-

rithm selector: selects the best ML algorithm based

on the workload pattern that is identified by the Con-

text Analyser. The selection of algorithms will be

based on the experiments that we performed to select

the best ML algorithm for each pattern. In other

words, it uses either SVR or NN for predicting the

future workload based on the day’s part.

Figure 6: Context-aware prediction framework.

7 CONCLUSION AND FUTURE

WORK

This paper has presented a CAPF to support auto-

scaling decisions in EC environments. This frame-

work predicts the future EC workload using either

SVR or NN ML algorithms. These algorithms are

considered the best ML algorithms to be used for EC

workloads which is based on the comparison that has

been performed. Further, the comparison process in-

cludes a thorough investigation on the window size

effect. All these steps have done using the Shanghai

Telecom dataset which represents a real EC work-

load. The results reveal that window size has a signif-

icant impact on workload patterns and ML algorithms

as best size allows the ML algorithms to capture the

workload pattern and behavior.

The SAS architecture is currently under develop-

ment with the aim of supporting, e.g. elasticity, scala-

bility, and QoS. In term of elasticity support, the pre-

dicted workload will be implemented to evaluate the

effectiveness of the developed models in operational

environment under several applications’ scenarios.

Further, this involves evaluating the performance of

the proposed SAS in the EC with consideration to var-

ious adaptation approaches which are proactive, reac-

tive, and hybrid adaptation. The scalability support

must also be considered to efficiently utilize the EC re-

source and maximize the number of running applica-

tions in the EC environment in sense that the EC have

limited resources. It requires meeting the applications’

QoS of IoT devices which have very sensitive require-

ments such as low latency. The support of scalability

and QoS will also involve the consideration of their

trade-offs, which is key in service provision.

REFERENCES

Ajila, S. A., and Bankole, A. A. (2013). Cloud client

prediction models using machine learning techniques.

International Computer Software and Applications

Conference, 134–142.

Al-Dhuraibi, Y., Paraiso, F., Djarallah, N., and Merle, P.

(2018). Elasticity in Cloud Computing: State of the Art

and Research Challenges. IEEE Transactions on

Services Computing, 11(2), 430–447.

Aldossary, M., and Djemame, K. (2018). Performance and

energy-based cost prediction of virtual machines auto-

scaling in clouds. 44th Euromicro Conference on

Software Engineering and Advanced Applications,

502–509.

Aljulayfi, A. F., and Djemame, K. (2019). A Novel QoS

and Energy-aware Self-adaptive System Architecture

for Efficient Resource Management in an Edge

Computing Environment. 35th Annual UK

Performance Engineering Workshop, 39–54.

Amiri, M., and Mohammad-khanli, L. (2017). Survey on

prediction models of applications for resources

provisioning in cloud. Journal of Network and

Computer Applications, 82, 93–113.

Arcaini, P., Riccobene, E., and Scandurra, P. (2015).

Modeling and Analyzing MAPE-K Feedback Loops for

Self-Adaptation. 10th International Symposium on

Software Engineering for Adaptive and Self-Managing

Systems, 13–23.

Baig, S., Iqbal, W., Lluis, J., and Carrera, D. (2020).

Adaptive sliding windows for improved estimation of

data center resource utilization. Future Generation

Computer Systems, 104, 212–224.

Bankole, A. A., and Ajila, S. A. (2013). Cloud client

prediction models for cloud resource provisioning in a

multitier web application environment. 2013 IEEE 7th

International Symposium on Service-Oriented System

Engineering, SOSE 2013, 156–161.

Calheiros, R. N., Masoumi, E., Ranjan, R., and Buyya, R.

(2015). Workload prediction using ARIMA model and

its impact on cloud applications’ QoS. IEEE

Transactions on Cloud Computing, 3(4), 449–458.

D’Angelo, M. (2018). Decentralized self-adaptive

computing at the edge. IEEE/ACM 13th International

Symposium on Software Engineering for Adaptive and

Self-Managing Systems, 144–148.

Delicato, F. C., Pires, P. F., and Batista, T. (2017). Resource

Management for Internet of Things. Springer.

https://doi.org/10.1007/978-3-319-54247-8

Dolui, K., and Datta, S. K. (2017). Comparison of edge

computing implementations: Fog computing, cloudlet

and mobile edge computing. 2017 Global Internet of

Things Summit (GIoTS), 1–6.

Framingham, M. (2019). The Growth in Connected IoT

Devices. IDC Analyze the Future. https://www.idc.

com/getdoc.jsp?containerId=prUS45213219

Galante, G., and De Bona, L. C. E. (2012). A survey on

cloud computing elasticity. 2012 IEEE/ACM Fifth

International Conference on Utility and Cloud

Computing A

, 263–270.

A Machine Learning based Context-aware Prediction Framework for Edge Computing Environments

149

Guo, Y., Wang, S., Zhou, A., Xu, J., Yuan, J., and Hsu, C.

H. (2019). User allocation-aware edge cloud placement

in mobile edge computing. Software - Practice and

Experience, January 2019, 1–14.

Gupta, H., Dastjerdi, A. V., Ghosh, S. K., and Buyya, R.

(2017). iFogSim: A toolkit for modeling and simulation

of resource management techniques in the Internet of

Things, Edge and Fog computing environments.

Software - Practice and Experience, 47(9), 1275–1296.

Islam, S., Keung, J., Lee, K., and Liu, A. (2012). Empirical

prediction models for adaptive resource provisioning in

the cloud. Future Generation Computer Systems, 28(1),

155–162.

James, G., Witten, D., Hastie, T., and Tibshirani, R. (2017).

Introduction to Statistical Learning with Applications

in R. Springer.

Kaur, K., Dhand, T., Kumar, N., and Zeadally, S. (2017).

Container-as-a-Service at the Edge: Trade-off between

Energy Efficiency and Service Availability at Fog Nano

Data Centers. IEEE Wireless Communications, 24(3),

48–56.

Kavanagh, R., Djemame, K., Ejarque, J., Badia, R. M., and

Garcia-perez, D. (2019). Energy-aware Self-Adaptation

for Application Execution on Heterogeneous Parallel

Architectures. IEEE Transactions on Sustainable

Computing, 1–15.

Kramer, J., and Magee, J. (2007). Self-Managed Systems:

an Architectural Challenge. 2007 Future of Software

Engineering, 259–268.

Krupitzer, C., Roth, F. M., VanSyckel, S., Schiele, G., and

Becker, C. (2015). A survey on engineering approaches

for self-adaptive systems. Pervasive and Mobile

Computing, 17, 184–206.

Kumar, J., and Singh, A. K. (2018). Workload prediction in

cloud using artificial neural network and adaptive

differential evolution. Future Generation Computer

Systems, 81, 41–52.

Li, G., Song, J., Wu, J., and Wang, J. (2018). Method of

Resource Estimation Based on QoS in Edge

Computing. Wireless Communications and Mobile

Computing, 2018.

Liu, B., Guo, J., Li, C., and Luo, Y. (2020). Workload

forecasting based elastic resource management in edge

cloud. Computers and Industrial Engineering,

139(0360–8352), 1–12.

Liu, C., Liu, C., Shang, Y., Chen, S., Cheng, B., and Chen,

J. (2017). An adaptive prediction approach based on

workload pattern discrimination in the cloud. Journal

of Network and Computer Applications, 80, 35–44.

Lorido-Botran, T., Miguel-Alonso, J., and Lozano, J. A.

(2014). A Review of Auto-scaling Techniques for

Elastic Applications in Cloud Environments. Journal of

Grid Computing, 12(4), 559–592.

Moreno-vozmediano, R., Montero, R. S., Huedo, E., and

Llorente, I. M. (2019). Efficient resource provisioning

for elastic Cloud services based on machine learning

techniques. Journal of Cloud ComputingAdvances,

Systems and Applications, 8(1).

Nikravesh, A. Y., Ajila, S. A., and Lung, C. H. (2017). An

autonomic prediction suite for cloud resource

provisioning. Journal of Cloud Computing, 6(1).

Nikravesh, A. Y., Ajila, S. A., and Lung, C. H. (2015a).

Evaluating Sensitivity of Auto-scaling Decisions in an

Environment with Different Workload Patterns. IEEE

39th Annual International Computers, Software &

Applications Conference, 415–420.

Nikravesh, A. Y., Ajila, S. A., and Lung, C. H. (2014).

Measuring prediction sensitivity of a cloud auto-scaling

system. Proceedings - IEEE 38th Annual International

Computers, Software and Applications Conference

Workshops, COMPSACW 2014, 690–695.

Nikravesh, A. Y., Ajila, S. A., and Lung, C. H. (2015b).

Towards an Autonomic Auto-scaling Prediction

System for Cloud Resource Provisioning. Proceedings

- 10th International Symposium on Software

Engineering for Adaptive and Self-Managing Systems,

SEAMS 2015, 35–45.

Sapankevych, N. I., and Sankar, R. (2009). Using Support

Vector Machines: A Survey. IEEE Computational

Intelligence Magazine, 2, 24–38.

Sguangwang.com. (2018). The Telecom Dataset (Shanghai

Telecom). http://sguangwang.com/TelecomDataset.html

Shi, W., and Dustdar, S. (2016). The Promise of Edge

Computing. Computer, 49(5), 78–81.

Singh, S., and Chana, I. (2015). QoS-Aware Autonomic

Resource Management in Cloud Computing: A

Systematic Review. ACM Computing Surveys, 48(3),

1–46.

Toczé, K., and Nadjm-Tehrani, S. (2018). A Taxonomy for

Management and Optimization of Multiple Resources

in Edge Computing. Wireless Commu. and Mobile

Computing, 2018, 1–20.

Wang, S., Guo, Y., Zhang, N., Yang, P., Zhou, A., and

Shen, X. S. (2019). Delay-aware Microservice

Coordination in Mobile Edge Computing: A

Reinforcement Learning Approach. IEEE Transactions

on Mobile Computing, 1–1.

Wang, S., Zhao, Y., Huang, L., Xu, J., and Hsu, C. H.

(2019). QoS prediction for service recommendations in

mobile edge computing. Journal of Parallel and

Distributed Computing, 127, 134–144.

Wang, S., Zhao, Y., Xu, J., Yuan, J., and Hsu, C. H. (2019).

Edge server placement in mobile edge computing.

Journal of Parallel and Distributed Computing, 127,

160–168.

Xu, M., and Buyya, R. (2019). Brownout Approach for

Adaptive Management of Resources and Applications

in Cloud Computing Systems. ACM Computing

Surveys, 52(1), 1–27.

Zhang, G., Eddy Patuwo, B., and Y. Hu, M. (1998).

Forecasting with artificial neural networks: The state of

the art. International Journal of Forecasting, 14(1), 35–

62.

CLOSER 2021 - 11th International Conference on Cloud Computing and Services Science

150