EWDLL Software Aging State Identification Based On LightGBM-LR Hybrid Model

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

2022 IEEE 22nd International Conference on Software Quality, Reliability and Security (QRS)

EWDLL: Software Aging State Identification based on LightGBM-LR Hybrid Model

Xueyong Tan and Jing Liu∗


2022 IEEE 22nd International Conference on Software Quality, Reliability and Security (QRS) | 978-1-6654-7704-8/22/$31.00 ©2022 IEEE | DOI: 10.1109/QRS57517.2022.00117

College of Computer Science, Inner Mongolia University, Hohhot, China


cstxy@mail.imu.edu.cn, liujing@imu.edu.cn
*corresponding author

Abstract—Android systems are prone to software aging due to rejuvenation operations will bring additional system overhead,
the accumulation of numerical errors and storage-related bugs so the core of software aging and rejuvenation research is
during long-term operation, resulting in gradual performance to accurately identify the software aging state. It improves
degradation and sudden system hang-ups. Thus, it is very critical
to accurately identify the aging state for improving the running the accuracy of identifying the aging state of the software.
reliability of Android systems. In this paper, we propose a novel Then, software rejuvenation is carried out in time to reduce
software aging state identification method, named EWDLL. It operational risks and costs, and to improve system reliability
first introduces the exponential Weibull distribution to simulate and availability. In the scenario for Android systems, accu-
the aging state transfer process of the Android system, then it rately detecting and identifying the occurrence of Android
uses Fuzzy Analytical Hierarchy Process (FAHP) to weight the
model parameters and resource utilization parameters. Finally, software aging is of great significance for improving the
the weighted dataset is fed into the LightGBM-LR model to system’s reliability and user experience. However, with the
identify the software state. The experimental results show that continuous development of the Android operating system and
our EWDLL method performs better in identifying the software the more complex application types, it becomes more difficult
aging state for Android system, i.e., it is 0.86% to 1.09% higher to identify the software aging phenomenon under the Android
in identification accuracy than the pure LightGBM-LR model,
about 10.00% and 4.54% to 4.95% higher than the traditional system.
models KNN and RF, and 1.97% to 3.09% higher than single The commonly used software aging identification methods
LightGBM model. Compared with the LR model, it has a include model-based, measurement-based, and hybrid meth-
maximum accuracy improvement of about 33.29% to 35.64%. ods. The model-based method usually uses Petri nets [3],
Markov models [4], Stochastic Processes [5], Probability The-
Keywords—software aging state identification; LightGBM-LR
model; exponential Weibull distribution; Android system ory [6] and other mathematical tools to establish the system
state transition model according to the running state of the
software system, and then calculates the occurrence probability
I. I NTRODUCTION
of software aging state to arrange the rejuvenation operation
During the long-term use of an Android device, it tends of the software. It is difficult to establish an accurate state
to experience response delays, performance degradation and transition model for large software systems, and as the size of
sudden hang-ups, which may be caused by aging software. the software system increases, the number of states constructed
These problems greatly degrade the user experience, so it in this way grows exponentially, which leads to an explosion of
is necessary to conduct software aging studies on Android the state space [7]. At present, the software aging identification
systems, thereby improving the reliability of the Android methods based on measurement mainly select the historical
system. Software aging refers to the phenomenon of system data of system performance as aging indicators, such as CPU
performance degradation, system shutdown, or even complete usage, free memory, swap memory, etc., and then set the
software failure in a long-running software system. Software system aging threshold, and use time series analysis, machine
aging occurs due to the lack of allocable resources, severe learning and other methods to predict the changing trend of the
fragmentation of software systems, and the long-term accumu- historical data of aging indicators, and software rejuvenation
lation of internal errors such as memory leaks and improperly operation is performed before the predicted value reaches the
released threads, which are difficult to detect and eliminate aging threshold. It was found that model-based approaches
during the software development and testing phases [1]. The tend to have lower identification accuracy due to their lack
Android system itself has a complex architecture, and the of flexibility as they ignore the real operating conditions of
application and release of computing resources in the system the system. The measurement-based method is not suitable
are relatively frequent. Especially when performing high-load for wide application because of its different adaptability to
and high-power applications such as live video streaming different software. The hybrid method based on model and
and mobile games, software aging is prone to occur, which measurement can better deal with these problems. For exam-
seriously affects the user experience. ple, the hybrid model based on Artificial Neural Network and
To avoid software aging, Huang et al. proposed software Autoregressive Integrated Moving Average (ARIMA) model
rejuvenation. It is a proactive and preventive fault-tolerant [8], and the hybrid model based on Long Short-Term Memory
technology that cleans the system operating environment and (LSTM) model and ARIMA model [9], the experimental
resets the internal state of the system to prevent major failures results show that the identification accuracy of the hybrid
of the system in the future [2]. However, frequent software model is higher than that of a single model. However, these

2693-9177/22/$31.00 ©2022 IEEE 1146


DOI 10.1109/QRS57517.2022.00117

Authorized licensed use limited to: Indian Institute of Technology (ISM) Dhanbad. Downloaded on April 03,2023 at 11:03:03 UTC from IEEE Xplore. Restrictions apply.
methods all need to define the system state by artificially resident memory as the indexes to observe software aging,
dividing the boundary, which is not universal in the actual and proposed that the virtual memory and resident memory
operating environment. will gradually decrease during the long operation of Android
In this paper, we propose a novel software aging state system. Y. Qiao et al. showed that the software aging caused
identification method, named EWDLL, based on exponential by memory leakage is reflected in the perspective of users.
Weibull distribution (EW) and the Light Gradient Boosting Memory leakage in different heap areas has different perfor-
Machine-Logistic Regression (LightGBM-LR) for software mances through injection memory leakage [12].
aging state identification of Android system. Our method Model-based methods and measurement-based methods are
includes the following three highly related parts. First, we in- also prevalent in the study of software aging and rejuvenation
troduce EW distribution to simulate various influencing factors in mobile systems. Model-based methods construct system
in the software aging state transfer process, and then introduce state transition models with probabilistic characteristics to
the Fuzzy Analytic Hierarchy Process (FAHP) to weight the represent the software aging process. C. Weng et al. used
model parameters and resource utilization parameters. At last, a linear regression model to analyze the usage of available
the LightGBM model is used as the learning layer [10]. The memory under time series and used Markov chains to classify
processing result of LightGBM for the dataset is input to LR the Android system into four states: initial state, healthy state,
as the feature set, and the identification operation is executed sub-healthy state, and aging state. It was also pointed out
at the specific state transfer time point calculated by the EW that mild rejuvenation could not effectively mitigate software
distribution. The final recognition result of Android system aging [13]. J. Xiang et al. modeled the usage behavior and
software aging is obtained. When the system is identified as aging process by stochastic Petri nets, then combined them
being in an aging state, reasonable rejuvenation operations into a continuous-time Markov chain and proposed an active
can be performed in time, such as restarting the system to rejuvenation strategy based on the model where the Android
restore the system to the robust state, thus greatly reducing system was in sleep mode and restarted before the aging state
the shutdown of the Android system caused by software aging [14]. Y. Qiao et al. constructed a Markovian regeneration
and improving the reliability of the Android system. process for evaluating steady-state system availability and
The main contributions of this paper are as follows. optimizing the trigger time of regeneration strategies for the
• A software aging state identification method is well Android system [15].
constructed for the Android system, which integrates EW Although the model-based method is more generalized,
distribution, FAHP and LightGBM-LR. It uses the EW these models simplify the design of the system will af-
distribution to simulate the operating state of the Android fect the detection accuracy due to the constant iterations of
system to obtain the possible time and model parameters Android and hardware differences among mobile devices.
of software aging. While the FAHP is integrated to weight Measurement-based methods focus on the collection and anal-
the model parameters and resource utilization parameters. ysis of data from real operating environments. D. Cotroneo
Finally, the weighted data is fed into the LightGBM- et al. designed and wrote the Android aging detection and
LR model for classification. The classification model repair tool the ADaRTA, which detects the aging state and
improves the learning capability of the LR model by estimates the aging time by selectively monitoring the trends of
feeding the feature data processed by the LightGBM the departmental performance metrics through heuristic rules,
model into the LR model. and arranges and implements rejuvenation according to the
• We performed effective experiments to show that our expected aging time [16]. D. Cotroneo et al. used the Mann
EWDLL method has higher identification accuracy than Kendall test and the Sen process system to analyze whether
the traditional models KNN and RF. In addition, the there is an aging trend in the time series of system resources,
hybrid model has indeed improved the identification ac- and pointed out that the software aging problem is due to
curacy compared to the single LightGBM and LR models. the specific processes and services within the Android system.
The remainder of the paper is organized as follows. Sec- The system User Interface process was found to consume a
tion II introduces the related work. Section III describes the lot of system resources through experiments, and low-quality
detailed process of establishing the EWDLL model. Section applications can easily lead to software aging, at the same
IV analyzes the experimental results. We conclude the paper time, it is pointed out that the performance degradation is
in the last section. related to the kernel resource utilization, and a series of kernel
resource utilization indexes are proposed to judge whether the
II. R ELATED W ORK current system is aging and perform rejuvenation operation
With the widespread use of mobile devices and the rapid accordingly. Finally, from the perspective of users, launch time
development of Android and its applications, there has been an is used to indicate software aging [17]. V. Yakovyna et al.
increasing amount of research into software aging for Android. proposed two user-perceived metrics for the Android software
Users will experience lag and program crashes when using aging, namely frame drawing time and number of Jankey
mobile devices for a long time, so in 2013, a related study frames, which the authors used to detect software aging trends
[11] firstly proposed that the software aging phenomenon in relation to traditional metrics such as launch time [18].
exists in the Android system, selected virtual memory and D. Cotroneo et al. proposed a rejuvenation strategy that uses

1147

Authorized licensed use limited to: Indian Institute of Technology (ISM) Dhanbad. Downloaded on April 03,2023 at 11:03:03 UTC from IEEE Xplore. Restrictions apply.
customized software for software rejuvenation by identifying aging identification, this will greatly improve the overall model
and cleaning up bloated data structures, reducing the cost execution and reduce the time required for rejuvenation.
of application or system reboots [19]. With the development 2) Parameter weighting: As identifying the aging state of
of machine learning and deep learning, software aging has software involves many model parameters and system resource
been predicted using machine learning algorithms and neural parameters, but these parameters are not the same for software
networks. C. Wang et al. used Random Forest (RF) algorithm aging identification. AHP method is widely used due to its
to classify system states in the area of the Android system characteristics of combining qualitative and quantitative pro-
[20]. S. Huo et al. used deep learning for the first time cessing of various evaluation factors, as well as the advantages
to predict launch time and page fault, and compared with of being systematic, flexible and concise, but the biggest
Support Vector Machine (SVM) and Decision Tree to obtain problem of AHP is that when the evaluation index reaches
better accuracy [21]. Y. Qiao et al. used the Long Short-Time 4, its consistency is difficult to achieve ensure. In this case,
Memory (LSTM) to predict Heap Memory aging metrics for FAHP is proposed, which combines the advantages of the
prediction and compared with traditional statistical models fuzzy method and AHP, which can effectively reduce excessive
such as the ARIMA model and Multilayer Perceptron to obtain subjectivity and ensure consistency. The FAHP is used to
better accuracy [22]. weight model parameters and the real-time hardware resource
Through the analysis of the related work, we found that the utilization in the dataset. The use of FAHP enhances the
software aging indicators of Android systems mainly include degree of control over the various factors in the software aging
memory, lack of pages and start-up time, which ignore the process.
situation that the system is performing tasks and the cost 3) Software aging status identification: We will use the
of software rejuvenation, so the selection of software aging model to classify the weighted data. The LR model is one
indicators needs to be added. In addition, the widely used of the most widely used classification models due to its
model-based and measurement-based methods for identifying simplicity, efficiency and ease of interpretation. However, the
the aging state of software are limited in their adaptability, LR model has the problem of low learning ability, so if
which affects the accuracy of the model identification. How- LightGBM is used as the learning layer and then the results
ever, hybrid models can effectively solve this problem, so we of its processing dataset are fed into the LR model as the
propose the EWDLL hybrid model to identify the software feature set, this can overcome the problem of low model
aging state of Android systems. learning ability of the LR model and effectively improve
the software aging state identification accuracy. We use the
III. C ONSTRUCTION OF THE EWDLL M ODEL LightGBM-LR model to obtain the software aging status of the
Android system, according to which we can arrange software
At present, the widely used software aging state identi- rejuvenation operations in time.
fication methods model-based and measurement-based have
the problem of low flexibility and adaptability. The hybrid A. State Transition Model of Software System.
method of model-based and measurement-based is proposed. The state transition model of the Android system is shown
This paper proposes the EWDLL hybrid model to identify in Figure 2. In the figure, r1 , r2 , r3 , r4 and r5 are the
the aging state of the Android system. Figure 1 shows the transition rates between states, state S0 represents the normal
construction flow chart of the EWDLL model. operation state, state S1 represents the aging state, state S2
As can be seen from Figure 1, the EWDLL model mainly represents the failure state, and state S3 represents the software
includes three steps. rejuvenation state. When the program starts to run, the system
1) Software state transition model construction: The is in the S0 state, and the state S0 is a state that will not fail.
Weibull distribution is one of the most widely used distri- After a period of operation, aging occurs with the extension
butions in reliability analysis and is often used to estimate of time, and it enters the S1 state that may fail. The use
important life characteristics of products. However, since the of rejuvenation technology can make the system enter the
Weibull distribution does not effectively address the case S3 state with a high probability in the S1 state, and enter
where the image of the failure rate function is of the bathtub the S2 state with a very small probability. The distribution
and inverted bathtub type, the EW distribution was proposed model for the system state transition mainly includes the
and its distribution function applies to more types of data. Gamma distribution, Weibull distribution and EW distribution.
The EW distribution is used to construct an aging state Table 1 compares the changing trends of the failure rate
transition model for the Android system. This transition model functions of the Gamma distribution, Weibull distribution and
allows the current state task completion rate to be obtained, EW distribution.
which is one of the important parameters used to identify the Considering that the EW distribution is more flexible than
aging state of the software. Since software aging identification the Weibull distribution and the Gamma distribution has a
requires time and model running costs, it is not necessary to more flexible failure function. Since the most popular two-
determine the software state at each unit point in time during parameter Weibull does not characterize the non-monotonic
the continuum of software system operation. Therefore, when failure rate well, the EW distribution is introduced to simulate
the state transition model reaches the aging state for software the influencing factors in the process of software aging state

1148

Authorized licensed use limited to: Indian Institute of Technology (ISM) Dhanbad. Downloaded on April 03,2023 at 11:03:03 UTC from IEEE Xplore. Restrictions apply.
Figure 1. Flow chart of construction of the EWDLL model

Table 1. Comparison of Function Failure Rates of Different Distribution.


Parameter Gamma distribution Weibull distribution EW distribution
α=γ=1 Constant Constant Constant
γ>1, increase from 0 to λ γ>1, increase from 0 to ∞ γ>1, increase from 0 to ∞
α=1
γ<1, decrease from ∞ to λ γ <1, decrease from ∞ to 0 γ<1, decrease from ∞ to 0
α>1, increase from 0 to λ α>1, increase from 0 to λ α>1, increase from 0 to ∞
γ=1
α<1, decrease from ∞ to λ α<1, decrease from ∞ to λ α<1, decrease from ∞ to 0
α,γ>1 increase from 0 to λ increase from 0 to ∞ increase from 0 to ∞
α,γ<1 decrease from ∞ to λ decrease from ∞ to 0 decrease from ∞ to 0

 α
x−λ
F (x) = 1 − exp −( ) , x ≥ λ, α, λ, γ > 0 (1)
γ
The corresponding probability density function is as fol-
lows.

 α  α−1
α x−λ x−λ
f (x) = × exp −( ) × (2)
γ γ γ
Where α is the first shape parameter, γ is the second shape
parameter, and λ is the scale parameter. This distribution
Figure 2. Software system state transition diagram can be recorded as EW (α, γ, λ), also known as the three-
parameter EW distribution.
The cumulative distribution function, also known as the
aging function, actually marks the probability that a software
system will age from 0 to t. The probability that a software
transition, and then combined with the real-time changes of system will not age is called the reliability function and is
resource utilization parameters to analyze the process changes expressed as follows.
of the software system more accurately [23]. Therefore, this  α
x−λ
paper uses the EW distribution to simulate the software aging r(x) = 1 − F (x) = exp −( ) (3)
of Android systems. The combination model is based on the γ
EW distribution, the distribution function of EW distribution The mathematical expression for the failure rate function is
is as follows. as follows.

1149

Authorized licensed use limited to: Indian Institute of Technology (ISM) Dhanbad. Downloaded on April 03,2023 at 11:03:03 UTC from IEEE Xplore. Restrictions apply.
 α−1 i
f (x) α x−λ
= ×
X
η (x) = (4) C= mi × ci (6)
r(x) γ γ
i=1
According to Figure 2, the Android system also includes
four states, the normal operation state, the aging state, the A state transition model for software aging is constructed
failure state, and the software rejuvenation state. The time through the EW distribution, enabling the simplification and
distribution function of mutual transformation between each abstraction of the complex practical problem of software
state is set as a three-parameter EW distribution. aging into an analyzable mathematical problem. Based on this
In the aging state transition model of software systems, re- model, the next subsection will use FAHP to weight the model
juvenation cost, system reliability and system state are mainly parameters and resource utilization parameters as the input
considered. The rejuvenation cost is used to grasp the overall dataset for the LightGBM-LR model.
rejuvenation time cost, it cost mainly includes the computing
B. Parameter Weighting Based on the FAHP
and storage resources consumed by the rejuvenation. The
number of rejuvenation operations performed by the system Following the construction of an EW distribution-based
in a state is used to grasp the rejuvenation situation of the software system aging state transition model, in order to
system in each state, and the maximum rejuvenation times improve the reliability of the aging state identification of An-
of the system is used to limit the number of rejuvenation droid systems, the dataset consisting of the model parameters
operations performed and reduce rejuvenation costs as far as and real-time resource utilization parameters was coefficient
possible. Especially for the system to perform key tasks, it is weighted before being identified. This enhances the subjective
particularly necessary to consider the reliability of the system and objective rationality of the whole process and its flexibility
and the probability of being able to complete the current task. in different application environments.
Since the time cost of performing rejuvenation operations in The steps of FAHP mainly include clarifying problems,
different states is different. The parameters involved in the establishing an analytic hierarchy process structure model,
model calculations are described in Table 2. constructing a judgment matrix, hierarchical single ranking
Table 2. Description of Model Parameters
and consistency test, hierarchical total ranking and consistency
test. FAHP was proposed to reduce the subjectivity of the AHP
Paramater Description model. The FAHP method is suitable for solving complex
T Maximum processing time allowed by the system
N Number of tasks operated by the system decision-making problems, especially in the evaluation of
M Maximum number of rejuvenation sessions multi-index and fuzzy scheme optimization, which can obtain
ki System task processing speed in state i more satisfactory optimization results. There are a lot of
mi Number of rejuvenation performed by the system in state i
ci Rejuvenation cost of the system in state i factors in the identification of Android system software aging,
cu Rejuvenation unit cost so we choose FAHP to select the key factors for software aging
C Total rejuvenation cost identification.
The FAHP method needs to determine an appropriate
Based on the software system aging state transition model, decision-making problem as the target problem to be analyzed
this paper makes the following task settings. The number of from the perspective of determining the target decision-making
task operations to be completed by the software system is set problem [24]. The problem must contain multiple influencing
to N . Due to the real-time requirements of the task, this paper factors to analyze the composition of the influencing factors
sets the maximum processing time allowed by the system to of the target problem and determine the relationship between
T . Once the task is not completed after this time, it will various factors. Secondly, based on the constituent units of
be judged as a task failure. In the rejuvenation operation, each decision-making influencing factor, the multi-level struc-
as the rejuvenation operation will cause a certain degree of ture model of the decision-making problem is determined. In
time loss, for the convenience of calculation, this paper sets this step, a decision-making problem is transformed into a
several parameter values. The total rejuvenation cost C is hierarchical decision-making system, which is convenient for
used to capture the overall rejuvenation time cost, the number decision-makers to determine the weight of each influencing
of rejuvenations performed by the system in state i is used factor. We then need to build a fuzzy judgment matrix to
to capture the rejuvenation situation of the system in each determine the relative weights by comparing them with each
state, and the maximum number of rejuvenation operations other and to determine the weight ranking of each influencing
M is used to limit the number of rejuvenation operations. The factor in the overall model. Then, the weights of each influ-
formula for calculating the maximum number of rejuvenations encing factor in the same layer are determined according to
in the current state is as follows. the construction of the judgment matrix, and then calculate
the specific weight of each influence factor in the model.
T − N × ki Finally, the iterative accuracy of the index weight needs to be
M= (5)
cu processed through the inverse comparison matrix. By judging
C represents the total rejuvenation cost and is calculated as whether the iterative weight error meets the accuracy require-
follows. ments of the iterative termination conditions, we can judge

1150

Authorized licensed use limited to: Indian Institute of Technology (ISM) Dhanbad. Downloaded on April 03,2023 at 11:03:03 UTC from IEEE Xplore. Restrictions apply.
Figure 3. The FAHP construction process

whether it is necessary to continue the iterative adjustment of the rejuvenation operation is executed at this time and whether
the judgment matrix until the termination conditions are met, the system can complete the task within the specified time.
and then the final convergence result can be obtained. The Therefore, the probability of completing the task in the current
flow chart of the FAHP method is shown in Figure 3. state is introduced. Finally, related research shows that real-
After determining the decision-making problem that needs time hardware resources also have a partial impact on the
to be analyzed, the analysis target is placed at the top level of software state, and it also reflects the running state of the
the hierarchical structure analysis model. This layer is called software system to a certain extent, so this paper introduces
the target layer, which means the ultimate goal of solving the the resource utilization parameter [25].
decision-making problem. Next, according to the established To quantitatively describe the relative importance of any two
problem analysis goal, it is necessary to build a hierarchical factors concerning a criterion in the upper level, this paper uses
structure analysis model based on the influencing factors of the 1 to 9 scale shown in Table 3 to give a quantitative scale.
the decision-making problem, and decompose the decision-
Table 3. Fuzzy Judgment Matrix 1 to 9 Scale
making problem to be analyzed into various components that
aij Description
affect its goal. 1 Equally important
3 Slightly important
5 Obviously important
7 Much more important
9 Extremely important
2,4,6,8 Between two adjacent levels of importance
1
aij
The importance ratio is a1
ij

In this paper, according to the system reliability is more


important than the rejuvenation cost, and the rejuvenation cost
is more important than resource utilization, the obtained first-
level judgment matrix is as follows.
1 3 13
 
 3 1 5  (7)
1 1
3 5 1
Figure 4. The FAHP software rejuvenation analysis structure diagram
The secondary decision matrix is as follows.
As shown in Figure 4, this paper sets the rejuvenation
decision as the target layer, the rejuvenation cost, system 1 1
 
1 3 7
reliability and system real-time status as the criterion layer, and  3 1 1  (8)
5
the maximum rejuvenation times in the current state, the task 7 5 1
completion probability in the current state and the resource
utilization as the scheme layer. This setting is based on the 1
 
1 3 5
following considerations, firstly, this paper needs to grasp the  3 1 5  (9)
rejuvenation situation in real-time and dynamically, so as to 1 1
1
5 5
control the rejuvenation cost, so the maximum rejuvenation
times in the current state are introduced. Secondly, the reli-
 
1 3 6
ability of the software system must be guaranteed, when the  1
1 5  (10)
5
software system makes a decision, it should consider whether 1 1
1
6 5

1151

Authorized licensed use limited to: Indian Institute of Technology (ISM) Dhanbad. Downloaded on April 03,2023 at 11:03:03 UTC from IEEE Xplore. Restrictions apply.
To avoid judgment logic errors in the judgment matrix, a identify software aging states, which is the core research idea
logical consistency check should be carried out. The above of this paper.
matrices all meet the consistency test conditions, and then five The LightGBM model and the LR model have obvious
eigenvectors are obtained through calculation. advantages in their respective fields, but their disadvantages are
    also very obvious. LR is a generalized linear model, which is
0.258 0.637 0.067 different from other linear models because it uses the Sigmoid
 0.618   0.297   0.007  (11) function so that the final function maps within [0,1]. It is
0.297 0.618 0.055 widely used in the field of commercial computing and has
   achieved relatively good results. Although the computational
0.030 0.008 complexity of the LR model is not high and the parallelized
 0.016   0.062  (12) data processing performance is excellent, due to the limitation
0.025 0.005 of its linear model structure, its model learning ability is
The judgment matrix is the basic information of the FAHP, not high, and a certain scale of feature analysis engineering
and it is an important basis for calculating relative importance is often required before application. Feature engineering in
and performing hierarchical sorting. The judgment matrix the LR model is very important, and the effect of feature
allows us to obtain a ranking of the importance of each param- combination is very limited if relying on artificial experience.
eter for identifying the aging state of the software. According Therefore, how to improve the efficiency, scale and automation
to the above process, the final weighting coefficients of model of feature selection engineering have always been a problem
parameters and real-time resource utilization parameters based worthy of research. Including academia and industry, many
on FAHP are shown in Table 4. studies have tried to combine the two models of LightGBM
and LR, and they have performed well [26]. Figure 5 shows
Table 4. Model Parameter Weight Coefficients
the basic structure of the LightGBM-LR hybrid model.
Paramater Coefficient
The maximum number of rejuvenation times 0.391
Current state task completion rate 0.517
CPU usage 0.024
Memory usage 0.043
Frame loss rate 0.025

C. Software Aging State Identification Based on the


LightGBM-LR Hybrid Model
While the parameters have been weighted by FAHP in
the previous section, the ultimate aim of this paper is to
accurately identify the aging state of the software. However,
considering the real-time requirements of software rejuvena-
tion, some adjustments have been made to the timing of state
classification. Due to its special application scenario, software Figure 5. The LightGBM-LR hybrid model structure
rejuvenation with high real-time requirements, software aging
state discriminating algorithms cannot work continuously in As shown in Figure 5, the LightGBM algorithm produces 2
continuous time intervals, which can neither meet the real-time decision trees with 5 leaf nodes, where the number 1 indicates
requirements of software rejuvenation, but also cause a con- that the predicted outcome of the training sample x falls on
siderable waste of time, hardware and computing resources. that leaf node. The predicted result can be expressed in the
As the EW distribution has been verified, it can well simulate One-Hot code as [0, 1, 0, 0, 1]. The leaf nodes of these trees
the software aging state transition process. After quantifying will then be used as feature vectors for LR, with the leaf node
the model parameters, preliminarily determine the time nodes on which the sample falls being 1 and the rest of the nodes
where the software state may change through calculation, being 0. With more dimensional tables for the change features
and judge the aging state on these time nodes calculated by above, the numerical data is also transformed into categorical
the model, which can not only greatly reduce the frequency data, which facilitates classification by the LR model.
used by the algorithm, but also theoretically improve the It can be seen from Figure 6 that the hybrid model uses the
accuracy of the algorithm. As model-based and measurement- LightGBM model as a learning layer and inputs the LightGBM
based methods were found to be inflexible and less versatile prediction result for the dataset as a feature set into the output
during the research process, this paper introduces the idea layer model LR. Finally, the output of the LR model is used as
of combining model and measurement methods to identify the output of the entire hybrid model. The process of training
software aging states. Considering the many factors involved the LightGBM-LR model is mainly divided into the following
in software aging, this paper chooses to use a combination of steps.
EW distribution, FAHP and LightGBM-LR model to jointly • Train the LightGBM model using the dataset.

1152

Authorized licensed use limited to: Indian Institute of Technology (ISM) Dhanbad. Downloaded on April 03,2023 at 11:03:03 UTC from IEEE Xplore. Restrictions apply.
Figure 6. The LightGBM-LR hybrid model flow chart

• Use the LightGBM model for predictive analysis as part research background, it can be seen that the operating state
of the feature learning project, One-Hot encodes the of the system and the user experience are very important to
output and feeds it into the LR model for training. the judgment of the real-time status of the system, and the
• The LR model will output the identification result of the combined index is more effective than a single index [27]. The
last software state. real data set reflects the system state change characteristics of
this Android device under the current operating environment
IV. E XPERIMENT AND R ESULT A NALYSIS
consisting of the number and types of Android applications.
This section mainly introduces the experimental results and The dataset is pre-processed by patching and cropping to
analysis of the software aging state identification based on the eliminate as much as possible the impact of software aging on
EWDLL hybrid model. Firstly, the experimental environment the data collection and to restore or preserve the system state
and the experimental dataset are presented. Secondly, the characteristics of the Android device in its current operating
results of FAHP weighting of model parameters and resource environment.
utilization parameters are shown. Thirdly, the results and In this paper, 10 sets of experiments were conducted to
analysis of the identification of parameters not weighted with simulate 10 days of mobile phone usage by users. Each set
FAHP using the LightGBM-LR model are presented. Finally, of experiments can show the Android system from the initial
the results and analysis of the identification of parameters state to the aging state, through 10 sets of experiments can
weighted by FAHP using the EWDLL model are described. effectively avoid the interference of external factors and ran-
A. Experimental Environment and Dataset dom errors in the experiment. Each experiment was conducted
to ensure that the Android device was in a healthy state,
The experimental part of the dataset collection platform
and then a load generator and system monitor were started
in this paper is the Android mobile terminal Mi 10, the
on the computer, where the Monkey stress test opened an
operating system is MIUI 11 based on Android 10, the CPU is
application 10,000 times and the system monitor recorded
Snapdragon 865, and the running memory is 8G. To investigate
system parameters and filtered out CPU usage, memory usage
the phenomenon of software aging in Android, the experiments
and frame loss rate at 30-second intervals.The Android perfor-
involve giving the phone a stress load so that the device enters
mance monitor is an application that uses the Dumpsys tool
the aging state as quickly as possible. This is followed by the
to get system parameters. In this paper, the Android Debug
examination of the phone’s operating parameters under stress
Bridge is used to connect the mobile phone to the computer.
load, which form a time series that is an important dataset for
Due to the randomness of the simulation of user behavior in
this paper to study software aging. The experiments are based
the experiment, the time from robust to aging of the Android
on the Monkey stress load tool and Dumpsys debugging tool
system in each group of experiments is different, about 4 to 8
to write load generation and performance monitoring scripts,
hours. Therefore, the amount of data collected in each group
and this paper uses Android service components to monitor
of experiments is inconsistent, from about 480 to 864 pieces,
system running parameters.
about 7,100 pieces in total.
In order to better verify the system running status and
user experience, the system monitoring tool is used to obtain
B. Weighting of Parameters Using the FAHP
samples of Android system performance metrics consisting
of CPU usage, memory usage and frame loss rate, which In this paper, Negative Log-Likelihood (NLL), Akaike In-
are stitched together in order to form a time series, with formation Criterion (AIC) and Bayesian Information Criterion
multiple time series forming the real dataset. The joint aging (BIC) are used as the standard for model comparison, accord-
indicator starts from two perspectives of system resources and ing to the likelihood ratio test to compare the goodness of
user experience, CPU usage and memory usage represent the fit between the EW distribution and the Weibull distribution.
usage of system resources, and the usage of system resources NLL is the value of the minimized negative log-likelihood
will directly affect the running status of the system, frame function, which can be used to compare the pros and cons of
loss rate represents the usage of user experience. From the models with the same parameters. The smaller the NLL value,

1153

Authorized licensed use limited to: Indian Institute of Technology (ISM) Dhanbad. Downloaded on April 03,2023 at 11:03:03 UTC from IEEE Xplore. Restrictions apply.
the better the model fitting effect. NLL can be expressed as of the Android system and then label the dataset. The data
follows. is then pre-processed to deal with missing and unusually
N LL = −L(θ) (13) incorrect data. The original dataset was then divided into a
training set and a test set according to 70% and 30%. The
The AIC test measures the goodness of fit of a statistical
final model is tested.
model. Its advantage is that it cannot only compare the pros
The LightGBM-LR classification model proposed in this
and cons of nested models and non-nested models, but also
paper is data-driven, which can be trained by a classifier
compare the pros and cons of different parameter models. Its
with a sufficient amount of historical labelled data to obtain
expression is as follows, where k is the number of model
an accurate classifier model. The main steps for conducting
fitting parameters.
experiments based on the LightGBM-LR classification model
AIC = 2k − 2L(θ) (14) are as follows.
1) Acquisition of data: A period of data from the same
BIC can compare the pros and cons of all models. The historical period needs to be selected from the database for
smaller the value of BIC, the higher the fitting degree of the the training and testing of the classification model.
model. Its definition expression is as follows. 2) Data pre-processing: The collected data needs to be
BIC = k ln(n) − 2L(θ) (15) processed to improve the quality and availability of the data.
As the raw data collected has problems such as missing,
In this paper, the Maximum Likelihood Estimation method redundant and containing noisy data, the data is pre-processed
is used to estimate the parameters of the EW distribution and to avoid these problems affecting the classification accuracy
the Weibull distribution, and NLL, AIC and BIC are used as of the model. The data pre-processing mainly includes data
the standard to compare the pros and cons of the models. The cleaning and dataset labeling, with the main focus on merging
model estimation results based on the Weibull distribution and multiple datasets collected and removing redundant data, and
EW distribution are shown in Table 5. manually labeling all datasets as aged or non-aged according
Table 5. Model Parameter Estimation Results to the actual operating conditions of the Android system. The
Model Parameter estimation NLL AIC BIC labeling criteria for the aging of the software system in the
α=0.567 training set and test set are mainly based on the following
Weibull λ=17.902 3532.91 7678.36 7691.02 three points.
γ=1.334
α=0.581 • The system hangs up for more than 5 seconds.
EW λ=18.957 3518.52 7659.08 7677.18 • No response prompt from any software under special
γ=1.572
circumstances.
• There is no abnormal restart of the software under special

Table 5 shows that the NLL, AIC and BIC values of the circumstances.
exponential Weibull distribution are smaller than those of the One of the above three conditions is met in order to be labeled
Weibull distribution. Therefore, under the evaluation criteria as the aging of the software system.
of this paper, we consider that the EW distribution is better 3) Dividing the training set and test set: The collected and
than the Weibull distribution for simulation. pre-processed data was divided into a training set and a test
set in the ratio of 70% and 30%.
C. The LightGBM-LR Model Identification Results and Anal- 4) Model training: The untrained LightGBM-LR classifier
ysis is trained with the training dataset, the model parameters are
After constructing a software system aging state transition initialized, the training data is trained by the LightGBM model,
model using EW distribution and a weighted indicator dataset the leaf nodes of the decision tree generated in LightGBM
using FAHP, this subsection will process the optimized dataset are combined to generate the feature training set, and then
using a classification algorithm based on the LightGBM-LR input to the LR classifier model for classification training, thus
hybrid model for software system aging state identification. In completing the training of the LightGBM-LR classifier.
this step, this paper mixes LightGBM and LR models and uses 5) Model testing: The test set is mainly used to test the
the output of LightGBM as the input of LR to classify the data. generalization ability of the model, so that the correct rate of
The classification datasets are model parameters and resource the training set and the test set can reach a balance, which
utilization parameters. These resource utilization parameters can prevent overfitting and improve the generalization ability
are also the basis of the above experiment. The first is to verify of the model.
the classification accuracy of the LightGBM-LR model in the 6) The model is saved: After the classifier model is estab-
application scenario of software aging state identification. lished, it needs to be saved and called when it is used.
Classifier model building generally goes through several In this experiment, five groups of comparative experiments
processes such as data collection, data pre-processing, dividing will be conducted to compare LightGBM-LR with the other
the training and testing sets, model training, model testing and four groups of common classification algorithms and classi-
model preservation. Firstly, we classify the collected dataset fication algorithm combination models. The four algorithms
into aging and non-aging states according to the performance are the K-Nearest Neighbor algorithm (KNN), Random Forest

1154

Authorized licensed use limited to: Indian Institute of Technology (ISM) Dhanbad. Downloaded on April 03,2023 at 11:03:03 UTC from IEEE Xplore. Restrictions apply.
Figure 7. Comparison results of the LightGBM-LR identification experiments

Table 7. Improved Accuracy of the LightGBM-LR Model


(RF), LightGBM and LR. The classification and identification Compared to Other Models
performance judgment indexes used are Accuracy, Precision, Model Accuracy improvement values
Recall and F1-Score. The above evaluation methods are calcu- KNN 9.98%
lated as follows. The experimental results are shown in Table RF 4.54%
LightGBM 0.89%
6 and Figure 7. LR 35.64%
correct
Accuracy = (16)
total
to the KNN, RF and LightGBM models, with the largest
TP improvement of 35.64% compared to the LR model. These
P recision = (17)
TP + FP results fully prove the idea of this paper that mixing the
LightGBM with the LR model can effectively improve the
TP
Recall = (18) classification capability.
TP + FN
D. The EWDLL Model Identification Results and Analysis
2 × P recision × Recall
F 1 − Score = (19)
Based on the previous experiments, this section uses FAHP
P recision + Recall
to process the dataset, and weights the coefficients of model
parameters and resource utilization parameters, still using
Table 6. Comparison Results of the LightGBM-LR Identification Experiments
Accuracy, Precision, Recall and the F1-Score as the judgment
Model Accuracy Precision Recall F1-Score
KNN 0.834 0.825 0.848 0.830 index of classification performance. The experimental results
RF 0.881 0.896 0.901 0.896 are shown in Table 8, Table 9 and Figure 8.
LightGBM 0.913 0.907 0.916 0.910
LR 0.679 0.699 0.722 0.685 Table 8. Comparison Results of the LightGBM-LR Identification
LightGBM-LR 0.921 0.927 0.917 0.920 Experiments under the FAHP
Model Accuracy Precision Recall F1-Score
KNN 0.841 0.836 0.855 0.857
From the experimental results, it can be seen that the RF 0.889 0.914 0.912 0.926
accuracy of KNN, RF, LightGBM and LR classifiers are LightGBM 0.924 0.918 0.924 0.926
83.4%, 88.1%, 91.3% and 67.9%. They are lower than 92.1% LR 0.700 0.719 0.747 0.695
LightGBM-LR 0.933 0.935 0.927 0.929
of the LightGBM-LR classifier. And their Precision, Recall
and F1-Score are also lower than the LightGBM-LR classifier.
It can be seen that the classification performance when using It can be seen from Table 8 and Table 9 that after the
the LightGBM-LR hybrid model is better than the LightGBM introduction of EW distribution parameters and FAHP process-
classifier and the LR classifier. ing, the identification accuracy of KNN, RF, LightGBM, LR
From Table 7 we know that the LightGBM-LR hybrid model and LightGBM-LR models have all improved. After repeated
improves 9.98%, 4.54% and 0.89% respectively compared experiments, only for the EWDLL model proposed in this

1155

Authorized licensed use limited to: Indian Institute of Technology (ISM) Dhanbad. Downloaded on April 03,2023 at 11:03:03 UTC from IEEE Xplore. Restrictions apply.
Figure 8. Comparison results of the LightGBM-LR identification experiments under FAHP

Table 9. Changes in Accuracy of Each Classifier under the FAHP Table 10. Improved Accuracy of the EWDLL Model
Model Accuracy improvement range Compared to Other Models
KNN 0.83%-1.30% Model Accuracy improvement values
RF 1.00%-2.01% KNN 10.00%
LightGBM 1.09%-1.11% RF 4.95%
LR 2.00%-2.80% LightGBM 1.30%
LightGBM-LR 0.86%-1.09% LR 33.29%

compared with other models, especially compared to the


paper, the identification accuracy is improved by 0.86% to
single LR model by about 33.29%, which fully proves that
1.09% on average compared to the LightGBM-LR model.
LightGBM can effectively improve the learning effect of the
Therefore, the identification accuracy of the EWDLL model
LR model and greatly improve its identification accuracy.
proposed in this paper is the highest.
In summary, the EWDLL hybrid model proposed in this
Through experiments, we verify the excellent performance paper is 0.86% to 1.09% higher than the LightGBM-LR
of the LightGBM-LR hybrid model in the identification scene model, about 10.00% and 4.54% to 4.95% higher than the
of the software aging state. We also compared a variety of traditional models KNN and RF, and about 1.97% to 3.09%
commonly used machine learning algorithms to identify the higher than the single model LightGBM. Compared with the
software aging state for the model parameters and resource LR model, it has a maximum accuracy improvement of about
utilization parameters datasets that did not use FAHP. The 33.29% to 35.64%.
experiments in this section use FAHP to process model
parameters and hardware parameter datasets to identify the V. C ONCLUSION AND F UTURE W ORK
software aging state. The experimental results show that the In order to better identify the aging state of the Android
identification accuracy of the LightGBM-LR hybrid model is system, the EWDLL method is proposed in this paper, based
better than other machine learning models whether or not on EW distribution and the LightGBM-LR hybrid model. By
FAHP is used to weight the dataset. Moreover, after using introducing EW distribution to simulate the transition process
FAHP, the identification accuracy of various machine learning of the software aging state, various influencing factors in the
models has been improved to varying degrees, which fully running process of the software system are quantified, and
proves that the EWDLL model proposed in this paper can then a more accurate process change analysis of the software
better identify the software aging state of the Android system. system is carried out in combination with real-time hardware
From Table 10 we can see that the EWDLL model improves parameter changes. Then, according to the application sce-
by about 10.00% and 4.95% respectively compared to the narios of different software systems, the FAHP is proposed,
traditional KNN and RF models, and by about 1.30% and which is also a pre-operation idea for processing datasets by
33.29% respectively compared to the single LightGBM and machine learning algorithms in different application situations,
LR models. The results show that the identification accuracy of which expands the dynamic concept of software aging. In
the EWDLL model proposed in this paper is indeed improved the final software aging state identification step, this paper

1156

Authorized licensed use limited to: Indian Institute of Technology (ISM) Dhanbad. Downloaded on April 03,2023 at 11:03:03 UTC from IEEE Xplore. Restrictions apply.
selects the hybrid classification algorithm of LightGBM-LR, [6] Z. Zheng, K. Trivedi, “Guest Editorial: Special Issue on Modeling and
which overcomes the problems of the low learning ability of Mitigation Techniques for Software Aging,” Software Quality Journal,
vol. 28, no. 1, pp. 3–5, 2020.
the LR model and the need for a certain scale of feature [7] K. Vaidyanathan, K. Trivedi. “A Comprehensive Model for Software
analysis engineering. In this paper, LightGBM is used as a Rejuvenation,” IEEE Transactions on Dependable and Secure Computing,
tool for feature combinations. The hybrid model uses the vol. 2, no. 2, pp. 124–137, 2005.
[8] Y. Yan, P. Guo and L. Liu, “A Novel Hybridization of Artificial Neural
LightGBM model as a learning layer, and the prediction results Networks and ARIMA Models for Forecasting Resource Consumption
of the LightGBM for the dataset are input into the LR as in an IIS Web Server,” in Proceedings of the 2014 IEEE International
a feature set. Combined with the FAHP weighted data set, Symposium on Software Reliability Engineering Workshops, 2014, pp.
437–442.
the identification operation is performed at the specific state [9] J. Liu, X. Tan and Y. Wang, “CSSAP: Software Aging Prediction for
transition time point calculated by the EW distribution. The Cloud Services Based on ARIMA-LSTM Hybrid Model,” in Proceedings
experimental results show that the EWDLL model proposed of the 2019 IEEE International Conference on Web Services (ICWS),
2019, pp. 283–290.
in this paper is 0.86% to 1.09% higher than LightGBM-LR [10] M. Qi, “LightGBM: A Highly Efficient Gradient Boosting Decision
in identifying the aging state of the Android system, about Tree.” in Proceedings of the Neural Information Processing Systems,
10.00% and 4.54% to 4.95% higher than the traditional models 2017, pp.3149–3157.
[11] J. Araujo, V. Alves, D. Oliveira, P. Dias, B. Silva, and P. Maciel. “An
KNN and RF, and 1.97% to 3.09% higher than the LightGBM. Investigative Approach to Software Aging in Android Applications,” in
Compared with the LR model, it has a maximum accuracy Proceedings of the IEEE International Conference on Systems, Man, and
improvement of about 33.29% to 35.64%. It is verified that Cybernetics (SMC). 2013, pp. 1229–1234.
[12] Y. Qiao, Z. Zheng, and F. Qin. “An Empirical Study of Software Aging
our EWDLL method has higher identification accuracy in the Manifestations in Android,” in Proceedings of the IEEE International
application field of software aging state identification. This is Symposium on Software Reliability Engineering Workshops (ISSREW),
conducive to more effective software aging and rejuvenation, 2016, pp. 84–90.
[13] C. Weng, J. Xiang, S. Xiong, D. Zhao, and C. Yang. “Analysis of
thus improving the reliability of the Android system. Software Aging in Android,” in Proceedings of the IEEE International
Although the EWDLL software aging state identification Symposium on Software Reliability Engineering Workshops, (ISSREW),
model proposed in this paper solves some problems existing 2016, pp. 78–83.
[14] J. Xiang, C. Weng, D. Zhao, A. Andrzejak and S. Xiong. “Software
in the field of software rejuvenation research, it still has Aging and Rejuvenation in Android: New Models and Metrics,” Software
shortcomings. Future research work will also focus on other Quality Journal, vol. 1, no. 11, pp. 1–22, 2019.
types of software systems other than mobile terminal software [15] Y. Qiao, Z. Zheng, Y. Fang, F. Qin and K. S. Trivedi. “Two-level
Rejuvenation for Android Smartphones and Its Optimization,” IEEE
systems. In addition, we will expand the aging model param- Transactions on Reliability, vol. 68, no. 2, pp. 633–652, 2019.
eters and relevant evaluation indicators of software systems [16] D. Cotroneo, L. De Simone, R. Natella and R. Pietrantuono. “A Config-
according to different software systems, so as to optimize the urable Software Aging Detection and Rejuvenation Agent for Android,” in
Proceedings of the IEEE International Symposium on Software Reliability
hierarchical design of the FAHP and make the processing of Engineering Workshops (ISSREW), 2019, pp. 239–245.
datasets more in line with the actual situation of different [17] D. Cotroneo, F. Fucci, A. K. Iannillo, R. Natella, and R. Pietrantuono.
software systems. “Software Aging Analysis of the Android Mobile OS,” in Proceedings
of the IEEE 27th International Symposium on Software Reliability
ACKNOWLEDGMENT Engineering (ISSRE), 2016, pp. 478–489.
[18] V. Yakovyna and B. Uhrynovskyi. “User-Perceived Response Metrics in
This work was supported in part by the National Natural Android OS for Software Aging detection,” in Proceedings of the IEEE
Science Foundation of China (No. 61662051), the Inner Mon- 15th International Conference on Computer Sciences and Information
golia Science and Technology Plan Project (No.2020GG0187), Technologies (CSIT), 2020, pp. 436–439.
[19] D. Cotroneo, L. De Simone, R. Natella, R. Pietrantuono and S. Russo.
the Engineering Research Center of Ecological Big Data, “Software micro-rejuvenation for Android mobile systems,” Journal of
the Ministry of Education, the Inner Mongolia Engineering Systems and Software, vol. 186, no. 1, pp. 1–21, 2022.
Laboratory for Cloud Computing and Service Software, and [20] C. Weng, D. Zhao, L. Lu, J. Xiang, C. Yang, and D. Li. “A Rejuvenation
Strategy in Android,” in Proceedings of the IEEE International Sympo-
the Inner Mongolia Key Laboratory of Social Computing and sium on Software Reliability Engineering Workshops (ISSREW), 2017,
Data Processing. pp. 273–279.
[21] S. Huo, D. Zhao, X. Liu, J. Xiang, Y. Zhong and H. Yu. “Using Machine
R EFERENCES Learning for Software Aging Detection in Android System,” in Proceed-
ings of the 10th International Conference on Advanced Computational
[1] M. Grottke, R. Matias, K. S. Trivedi, “The Fundamentals of Software Ag-
Intelligence (ICACI), 2018, pp. 741–746.
ing,” in Proceedings of 2008 IEEE International Conference on Software
[22] Y. Qiao, Z. Zheng, and Y. Fang. “An Empirical Study on Software Aging
Reliability Engineering Workshops, 2008, pp.1–6.
Indicators Prediction in Android Mobile,” in Proceedings of the IEEE
[2] Y. Huang, C. Kintala, N. Kolettis and N. D. Fulton, “Software Rejuvena-
International Symposium on Software Reliability Engineering Workshops
tion: Analysis, Module and Applications,” in Proceedings of the Twenty-
(ISSREW), 2018, pp. 271–277.
Fifth International Symposium on Fault-Tolerant Computing, 1995, pp.
[23] G. S. Mudholkar, D. K. Srivastava. “Exponentiated Weibull family for
381–390.
analyzing bathtub failure-rate data,” IEEE Transactions on Reliability, vol.
[3] E. Andrade, F. Machida, “Analysis of Software Aging Impacts on Plant
42, no. 2, pp. 299–302, 1993.
Anomaly Detection with Edge Computing,” in Proceedings of the IEEE
[24] Z. Xu and H. Liao, “Intuitionistic Fuzzy Analytic Hierarchy Process,”
International Symposium on Software Reliability Engineering Workshops,
IEEE Transactions on Fuzzy Systems, vol. 22, no. 4, pp. 749–761, 2014.
2019, pp. 204–210.
[25] V. Yakovyna, and B. Uhrynovskyi, “User-Perceived Response Metrics
[4] F. Machida, P.Maciel, “Markov Chains and Petri Nets for Software Reju-
in Android OS for Software Aging detection.” in Proceedings of the
venation Systems,” in Proceedings of the IEEE International Symposium
2020 IEEE 15th International Conference on Computer Sciences and
on Software Reliability Engineering Workshops, 2020, pp. 325–326.
[5] G. Levitin, L. Xing and Y. Xiang, “Cost Minimization of Real-time Mis- Information Technologies (CSIT), 2020, pp.436–439.
sion for Software Systems with Rejuvenation,” Reliability Engineering &
System Safety, vol. 193, no. 2, pp. 1–12, 2020.

1157

Authorized licensed use limited to: Indian Institute of Technology (ISM) Dhanbad. Downloaded on April 03,2023 at 11:03:03 UTC from IEEE Xplore. Restrictions apply.
[26] D. Wang, H. Xiao, Q. Sun and Y. Chen, “The Application of Deep [27] D. Cotroneo, F. Fucci, A. K. Iannillo, R. Natella, and R. Pietrantuono,
Learning Algorithm in Marketing Intelligence,” in Proceedings of the “Software Aging Analysis of the Android Mobile OS.” in Proceedings
2019 IEEE 3rd Advanced Information Management, Communicates, of the 2016 IEEE 27th International Symposium on Software Reliability
Electronic and Automation Control Conference (IMCEC), 2019, pp. Engineering (ISSRE), 2016, pp. 478–489.
1356–1360.

1158

Authorized licensed use limited to: Indian Institute of Technology (ISM) Dhanbad. Downloaded on April 03,2023 at 11:03:03 UTC from IEEE Xplore. Restrictions apply.

You might also like