Improving the Trade-Off between Performance and Energy Saving
in Mobile Devices through a Transparent Code Offloading Technique
R
ˆ
omulo Reis
1 a
, Paulo Souza
1 b
, Wagner Marques
1 c
, Tiago Ferreto
1 d
and F
´
abio Diniz Rossi
2 e
1
Polytechnic School, Pontifical Catholic University of Rio Grande do Sul
Ipiranga Avenue, 6681 - Building 32, Porto Alegre, Brazil
2
Federal Institute of Education, Science and Technology Farroupilha, Brazil
Keywords:
Mobile Cloud Computing, Code Offloading, Mobile Devices, Face Detection, Android.
Abstract:
The popularity of mobile devices has increased significantly, and nowadays they are used for the most diverse
purposes like accessing the Internet or helping on business matters. Such popularity emerged as a consequence
of the compatibility of these devices with a large variety of applications. However, the complexity of these
applications boosted the demand for computational resources on mobile devices. Code Offloading is a so-
lution that aims to mitigate this problem by reducing the use of resources and battery on mobile devices by
sending parts of applications to be processed in the cloud. In this sense, this paper presents an evaluation of
a transparent code offloading technique, where no modification in the application source code is required to
allow the smartphone to send parts of the application to be processed in the cloud. We used a face detection
application for the evaluation. Results showed the technique can improve applications performance in some
scenarios, achieving speed-up of 12x in the best case.
1 INTRODUCTION
The recent computing and telecommunication ad-
vances boosted the adoption of mobile devices. These
devices have been used for the most diverse purposes,
such as entertainment and personal care. However, as
the complexity of mobile applications increases, the
demand for computational resource increases as well.
So, there is a concern in supplying this increasing de-
mand without sacrificing the device’s batteries (Zhou
et al., 2015).
Currently, mobile devices use diverse software
and hardware technologies to deliver greater conve-
nience to their users. But, these devices still present
a poor performance in comparison to desktop com-
puters. This can affect the user experience when
some application is running resource-intensive tasks
like natural language processing. In addition, there
is a concern about the trade-off between application
a
https://orcid.org/0000-0001-9949-3797
b
https://orcid.org/0000-0003-4945-3329
c
https://orcid.org/0000-0003-3304-5611
d
https://orcid.org/0000-0001-8485-529X
e
https://orcid.org/0000-0002-2450-1024
performance, battery consumption, and temperature,
since increasing the computational capacity of these
devices implies higher temperature and battery con-
sumption.
In this context, Mobile Cloud Computing (MCC)
emerges with the proposal of delivering greater pro-
cessing power to mobile devices without adversely
affecting aspects in the temperature and the energy
consumption of these devices (Akherfi et al., 2016).
MCC brings to mobile devices the unlimited high pro-
cessing capacity provided by cloud computing. It also
makes code offloading techniques feasible. Code of-
floading allows parts of an application to be sent and
executed in the cloud. Once the code is processed in
the cloud, the result is sent back to the mobile device.
As a consequence of these advantages, several ap-
plications like Google Photos
1
and Apple Siri
2
have
adopted this technique (Sanaei et al., 2014).
Several code offloading techniques have been de-
veloped (Benedetto et al., 2017; Thakur and Verma,
2015). However, most of them require explicit
changes in the application source code, which limits
their use, since this code is rarely available to anyone.
1
Available at:<https://photos.google.com >.
2
Available at: <https://www.apple.com/ios/siri>.
Reis, R., Souza, P., Marques, W., Ferreto, T. and Rossi, F.
Improving the Trade-Off between Performance and Energy Saving in Mobile Devices through a Transparent Code Offloading Technique.
DOI: 10.5220/0007719703470354
In Proceedings of the 9th International Conference on Cloud Computing and Services Science (CLOSER 2019), pages 347-354
ISBN: 978-989-758-365-0
Copyright
c
2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
347
So, a transparent code offloading technique for An-
droid devices was proposed (de Oliveira et al., 2017)
to eliminate this limitation. This technique allows
the use of code offloading and does not require any
modification in the application code. Experiments
were conducted to prove the viability of this tech-
nique. However, the experiments were performed in
a local environment, using a local network to connect
the smartphone and a virtual machine, which is more
close to a mobile edge computing environment.
In this sense, this paper presents an evaluation of
this transparent code offloading technique in a real-
world cloud environment. The main objective was to
evaluate whether this technique was a feasible option
in a cloud computing environment and how efficient
it could be. So, we used the same face detection ap-
plication from (de Oliveira et al., 2017) to conducted
a set of experiments, but using virtual machine in-
stances from a public cloud provider (Google Cloud
Platform) instead of a local VM. A secondary objec-
tive was to provide an evaluation of the impact of In-
ternet latency when this code offloading technique is
adopted. So, we used a set of virtual machines with
heterogeneous capacities and from different regions
of the world.
The remainder of this paper is organized as fol-
lows: In Section 2 we present the concepts of Mo-
bile Cloud Computing and Code Offloading. In Sec-
tion 3, the architecture of the transparent code offload-
ing technique is explained. In Section 4 we present
the specifications and configurations of the experi-
ment environment. Section 5 is reserved for the ex-
planation and discussion of results. In the Section 6
we present the final considerations.
2 BACKGROUND
Cloud computing is defined as a paradigm that allows
access through the Internet to a shared set of config-
urable computational resources with unlimited capac-
ity. Such resources can be quickly provisioned and
released with minimal effort of management or inter-
action with the service provider. (Mell et al., 2011).
Mobile Cloud Computing (MCC) enabled the mobile
device to benefit from the unlimited resources pro-
vided by cloud computing (Thakur and Verma, 2015;
Shiraz et al., 2013).
Since mobile applications are becoming more
complex and supporting more and more functionali-
ties, more computational resources are required to run
those applications, which leads to greater battery con-
sumption. Therefore, the code offloading technique is
used to send parts of an application or even the en-
tire application to run in the cloud. This can improve
application performance, as well as extend the bat-
tery life of mobile devices (Thakur and Verma, 2015;
Benedetto et al., 2017; Chun et al., 2010; Qian and
Andresen, 2015; Flores et al., 2015).
There are several proposals for code offloading
techniques as an alternative to providing better per-
formance and energy consumption. Some of them re-
quire the developers to modify the application code
in order define the partitioning and migration of parts
from an application, so only the part with the great-
est demand for computational power is sent to the
cloud. Others approaches require full application mi-
gration to the cloud, but this implies communication
overhead, which can increase the execution time of
applications.
Unlike these related studies, De Oliveira et al.
(de Oliveira et al., 2017) proposed a transparent code
offloading technique for Android devices. They used
Xposed Framework to transparently modify Android
framework methods and send tasks to be executed re-
motely on a server. In this sense, the proposal does
not require any changes to the firmware source code
or Android system applications. To evaluate the pro-
posal, an Android application was developed with
the demand for computing resources using the native
methods of the Android operating system. The au-
thors emulated a local cloud using a traditional vir-
tual machine running Android-x86, which is accessed
by the mobile device over a wireless network. The
results showed that this technique can improve exe-
cution time for several application cases, as well as
minimize memory allocation.
3 ARCHITECTURE DESIGN
The environment assumed by the transparent code
offloading technique is composed of i) Android in-
stances of virtual machines running in a local work-
station or in the cloud and accessible to ii) physical
Android devices, which are connected to a network.
In this architecture, each mobile device has installed
the Xposed Framework, which can modify the behav-
ior of a set of Android Framework methods. Each
mobile device also has an Xposed Framework cus-
tom module enabled. This custom module has a set
of Android framework methods or application meth-
ods. When this module is enabled, it detects when a
specific method is requested and manipulates this re-
quest to run in the cloud.
Figure 1 illustrates an overview of the architec-
ture design, where an Android device is running two
applications: App A and App B; and computational
CLOSER 2019 - 9th International Conference on Cloud Computing and Services Science
348
Figure 1: Architecture design overview.
infrastructure running Android instances, which has
an application that plays the role of a service that pro-
cess tasks received from the devices. In this case,
Both applications are invoking some method from the
Android framework. However, App B is invoking a
method which was pre-defined to be offloaded in the
custom module. Therefore, just the method invoked
by App B is hooked by Xposed and sent to the cloud.
In the same context, each virtual machine runs a
service that is responsible for receiving and respond-
ing to requests from mobile devices. When an appli-
cation calls a method that demands more computing
resources and it is defined in the Xposed framework
module, the task is sent for processing in the cloud.
Otherwise, the method is executed locally. Therefore,
such architecture does not impact the operation of ap-
plications and methods that are not pre-defined in the
module.
By using this architecture, application developers
do not need to modify an existing application or cre-
ate an application that sends its methods to the cloud,
since the Xposed framework module is responsible
for this. In the same sense, if there is a third party ap-
plication that invokes the same method used by App
B, the method would be offloaded as well. Therefore,
the use of Xposed framework allows the code offload-
ing technique to be performed through a transparent
way to the application developer since the developer
of an application does not need to program the appli-
cation’s communication with the server.
The transparent code offloading technique is a so-
lution that allows the device benefits from cloud com-
puting resources and the resource code of the Android
or application is not available. At the same time, it is
agnostic to Android distribution, since they have the
same Android framework. Hooking Android native
method allows several applications to benefit from
code offloading. In addition, the device users can
choose when they want to use this code offloading
technique. This technique supports the implementa-
tion of decision algorithms to evaluate which part of
the application and when to offload this part. How-
ever, this is out of the scope of this paper, since our
goal is not to evaluate decision algorithms for code
offloading.
4 EVALUATION SCENARIO
We have conducted a set of experiments in order to
provide an evaluation of the transparent code offload-
ing technique in a cloud computing environment. The
environment for this evaluation has three main com-
ponents: (i) Face Detection Application, (ii) Applica-
tion Service Server, and (iii) Xposed Framework Cus-
tom Module.
The face detection application is the same from
De Oliveira et al. (de Oliveira et al., 2017). This ap-
plication uses the native Android framework method
FaceDetection to detect faces of people in an im-
age. We have chosen this method because it demands
high computational power. This application is imple-
mented as a regular application (there are no annota-
tions for offloading) to be executed in the smartphone.
Application service server runs inside a virtual
machine with Android-x86. This service applica-
tion handles offloading requests from mobile de-
vices. The Xposed Framework custom module con-
nects both components. It modifies the behavior of
the method detect(Frame frame) from class the vi-
sion.face.FaceDetector. Then, instead of executing
the code to detect faces in the picture, the smartphone
sends a request to the application service server and
waits for its response, in order to get back to the ap-
plication with the appropriate result. The images are
available in the storage of the VM instance.
The FaceDetector Builder is set in the accurate
mode in the application service server. This makes
detect(Frame) takes longer to be executed, however, it
should be more accurate, finding more faces than any
other mode available. Besides, the execution time of
the method on the VM is not supposed to take longer
than locally on the mobile device, since the cloud has
more processing power.
Figure 2 illustrates the sequence diagram of our
experiment. When the face detection application calls
FaceDetector.detect(Frame), the Xposed Framework
acts as a man in the middle, intercepting this method
invocation and executing the method from the module
we have developed instead of the original one. Then,
it sends the ID of the image to the server. Each image
has a predefined ID, enabling the application service
to identify the required image. This allows the ap-
plication server to handle different images. Then, the
Improving the Trade-Off between Performance and Energy Saving in Mobile Devices through a Transparent Code Offloading Technique
349
service application receives the image ID, creating a
Frame.Builder for the predefined image, which is lo-
cally available in the internal memory, and executes
detect(Frame) method. Finally, it sends back to the
smartphone the number of detected faces.
The method from the module receives the num-
ber of detected faces from the application server. This
allows the module to create a loop for receiving the
faces from the server. Right after the number of faces
is sent, the application service cast each face into a
string in JSON format using the Gson library and
sends them, one by one, to the smartphone. Upon re-
ceiving each face, the module has to extract the Face
object in JSON format from a string and append it in a
generic SparseArray. Once all the faces are received,
this SparseArray is returned to the application. We
have used JSON since it is not possible to make a Java
class serializable in this context.
Figure 2: Sequence Diagram of the Experiment.
The experiment data gathered was based on the
following metrics: time to establish the connection,
time to receive faces detected from the server, the time
taken by the server to detect faces, and memory usage.
For this evaluation, we selected 5 images with differ-
ent dimensions and number of faces. The images are
presented in Figure 3 and the images specifications
are available in Table 1.
Table 1: Images Specifications.
Name Width x Height (px) Size (KB) Number of Faces
Image 1 1024x768 1300 2
Image 2 1024x768 914 4
Image 3 1024x768 1600 5
Image 4 1024x768 674 5
Image 5 1024x768 1270 29
Table 2: Latency and average distance between the cloud
servers and the smartphone.
Cloud Region Latency
Approximate Distance from where
the experiments were conducted
Sao Paulo (Brazil) 25.7ms 12521km
Sydney (Australia) 406ms 1142.9 km
We used Ravello as an infrastructure provider. In
this sense, we created and instantiated a virtual ma-
chine with Android-x86 running the application ser-
vice, which performs face detection in the cloud. To
verify the impact of different computing resources
available on the virtual machines on the cloud, we
used the three virtual machines sizes described in Ta-
ble 3.
Internet latency can influence application perfor-
mance that uses code offloading techniques. There-
fore, we deployed the virtual machines in two differ-
ent regions (Sao Paulo and Sydney) to present the im-
pact caused by Internet latency regarding execution
time. Then, we verify the latency of the connection
with the servers in the different regions and present
the approximated distance between the servers and
from where the experiments were conducted. The re-
sults are presented in Table 2. For each image we ran
the experiment with different virtual machine sizes,
running in different regions. Table 4 presents the con-
figuration of each experiment.
In the mobile device was installed the devel-
oped application with a FaceDetection method, and
the Xposed framework. The framework modified
the method FaceDetection from class FaceDetector.
Then, through this application, instead of executing
the application to detect faces on the mobile device, it
sent a request to the server application to perform this
task.
The mobile device used in the experiments has the
following configuration: Android 5.1.1, CPU 4x 1.1
GHz Cortex-A7, 1GB of RAM, and 16GB of storage.
It is important to emphasize that the Xposed Frame-
work installation requires root access to the mobile
device, bootloader unlocked, and TWRP installed.
After the Xposed Framework installation, we in-
stalled our custom module and enabled it. Each exper-
iment was performed ten times for each configuration
presented in Table 4. The virtual machines were run-
ning the Android 6.0.1 from the open source project
Android-x86 with no modification in its source code.
In this virtual machine, we installed the application
service server for receiving the requests from the
smartphones and performs the face detection.
Table 3: Virtual Machines Specification.
Cloud VM Size
Specifications
Processor Memory Disk
Small 1vCPU 4GB 8GB
Medium 2vCPU 4GB 8GB
Large 4vCPU 4GB 8GB
5 RESULTS AND DISCUSSION
We analyzed three metrics to evaluate the effi-
ciency of the transparent code offloading technique
(de Oliveira et al., 2017). The first metric was the to-
tal time, that is the time taken by the face detection
CLOSER 2019 - 9th International Conference on Cloud Computing and Services Science
350
(a) Image 1 (b) Image 2 (c) Image 3
(d) Image 4 (e) Image 5
Figure 3: Images adopted in the experiments.
Table 4: Experiment Specifications.
Image Environment
Image 1 Smartphone
Cloud VM
Small Instance
Cloud VM
Medium Instance
Cloud VM
Large Instance
Sao Paulo Sydney Sao Paulo Sydney Sao Paulo Sydney
Image 2 Smartphone
Cloud VM
Small Instance
Cloud VM
Medium Instance
Cloud VM
Large Instance
Sao Paulo Sydney Sao Paulo Sydney Sao Paulo Sydney
Image 3 Smartphone
Cloud VM
Small Instance
Cloud VM
Medium Instance
Cloud VM
Large Instance
Sao Paulo Sydney Sao Paulo Sydney Sao Paulo Sydney
Image 4 Smartphone
Cloud VM
Small Instance
Cloud VM
Medium Instance
Cloud VM
Large Instance
Sao Paulo Sydney Sao Paulo Sydney Sao Paulo Sydney
Image 5 Smartphone
Cloud VM
Small Instance
Cloud VM
Medium Instance
Cloud VM
Large Instance
Sao Paulo Sydney Sao Paulo Sydney Sao Paulo Sydney
application on the smartphone to show the detected
faces. Total time includes the time taken to instanti-
ate the Java objects required by face detection method
and the communication time with the cloud servers in
the code offloading scenarios. The second metric we
analyzed was the time to detect faces. We consid-
ered this metric in order to evaluate the technique in
a more accurate way, being able to analyze the per-
formance impact brought by running the face detec-
tion method in the cloud, without considering external
factors like communication delay. The third metric
we analyzed during the experiments was the memory
usage, that is the amount of RAM consumed by the
smartphone. Given the fact that memory is one of the
hardware components responsible for delivering per-
formance to the smartphone users, we analyzed such
metric to verify the memory gains that offloading ap-
plications to the cloud could bring. Moreover, mem-
ory also affects the device power consumption, so re-
ducing the memory usage can also implicate in reduc-
ing the smartphone battery usage.
5.1 Overall Application Performance
As we can see in Figure 4 (a), using code offload-
ing technique and the virtual machine instance in Sao
Paulo increased the overall application performance
according to the virtual machine specification. The
best result in this region was achieved by using the
medium instance, where using code offloading im-
proved the application performance, when consider-
ing the total execution time. The medium size in-
stance resulted in a speed up of 4.8x, while the small
instance resulted in a speed up of 3.9x and 4.4x for the
large instance. In this sense, we can perceive that ex-
ecuting the face detection method in the large virtual
machine does not guarantee the best performance.
The face detection application relies on CPU pro-
cessing, and as a consequence, using 2 vCPUs instead
of 1 vCPU has improved the application performance
in more 92%. However, increasing the virtual ma-
chine capacity to 4 vCPUs led to a 15ms overhead.
This is because this application does not require too
much processing to take advantage of 4 vCPUs of the
large instance, so both the medium and the large in-
stances bought close results. Through such findings,
we can conclude that the medium instance has the
specifications that best fit the demand for resources
of the face detection application.
Executing the face detection method in the cloud
server located in Sydney (Figure 4 (b)) decreased
the overall application performance by approximately
344% due to the 406ms of communication delay be-
tween this cloud server and the smartphone (Table 2).
Improving the Trade-Off between Performance and Energy Saving in Mobile Devices through a Transparent Code Offloading Technique
351
5.2 Time to Detect Faces
The results of the time to detection faces also in-
dicated the viability of the use of code offloading
through the cloud. In all cases, the virtual machines
instances presented better results than the mobile de-
vice to run all application without using code offload-
ing. All results are illustrated in Figure 4 (c) and in
Figure 4 (d). The worse outcome was obtained per-
forming the code offloading with a small cloud in-
stance on Sydney with image 5. However, this result
also presented a speed up of 3.5x in execution time
if compared with the mobile device execution time
result. Such value was obtained since that virtual ma-
chine has less computing resources than the other in-
stances and the image tested present 29 faces to be
detected.
The best result was collected using image 1 when
was obtained performance speed up of 12x using a
cloud large instance. In the same sense, image 1
presents just two faces to be detected, and the large in-
stance presents more computing resources than small
and medium instances. Then, it provides better per-
formance during face detection by the cloud instance.
In addition, if we consider just the region of Sao
Paulo, the medium virtual machine presents the bet-
ter performance when used to detect faces of image 1,
where it was 1209% faster in the execution time.
The worst result was captured during the execu-
tion of image 5 using the small instance of the vir-
tual machine with a speed up of 4.7x, since the image
needs more computing resources to detect 29 faces.
Then, when the large instances were adopted, we ob-
tained a better execution time. The results also illus-
trate the impact of the number of faces in the images
in each instance of the virtual machine, where it is
provided better performance using large virtual ma-
chines with a small number of faces in the images.
5.3 Memory Usage
We also analyzed the memory allocated after the exe-
cution of the method detect. Figures 4 (e) and (f) illus-
trate the average RAM allocated and, as expected, less
memory was allocated when executing the method
in the cloud. We got RAM savings up to 55,68%
for image 4 using a small instance in Sydney. This
happens because, by running the method remotely,
it prevents the use of memory required to run the
method locally. This can lead making the battery last
longer. Considering the average smartphone memory
allocated for all experiments using cloud infrastrucure
and compared with the memory allocated running the
application only locally, the experiments has allocated
42,07% less memory for Sydney and 25,87% less
memory for Sao Paulo.
6 CONCLUSION
Oliveira et al. (de Oliveira et al., 2017) proposed an
approach for using code offloading where no modifi-
cation is required on the application source code nor
changes in the Android system image. It is trans-
parent and easy to use for Android device users and
developers. It also allows the user to enable or dis-
able the use of code offloading. However, the authors
have run their experiments in a local network using
virtual machines through VirtualBox. In this paper,
we brought the transparent code offloading technique
for Android devices to the cloud in order to verify the
viability of the proposed architecture using the cloud
infrastructure.
In this sense, we adopted Ravello as the cloud
provider and several metrics were collected. The first
metric was the time the application on the smartphone
needs to perform the face detection without the use
of cloud instances. We presented the evaluation of
the total time to perform the face detection using the
cloud, which includes aspects such as the time to in-
stantiate the face detection method and the time to
communicate with the cloud instances. In the same
context, we considered the time the applications took
to detect faces without considering external factors
like communication delay.
Two regions were considered and five images
were adopted during all experiments. Virtual ma-
chines with different computing resources were used
as well. Our results showed that this technique can
improve performance in terms of execution time and
minimize the RAM usage by using cloud comput-
ing. They also showed that this technique is highly
impacted by network infrastructure and Internet ac-
cess. Results indicated performance speed up of
12x% when using the cloud to perform code offload-
ing.
The resources from the cloud are more powerful
and all applications could theoretically use this tech-
nique. However, not all of them can benefit from this
technique, since not always running a task remotely is
faster than locally, keeping that in mind for improving
performance, the methods from the Android Frame-
work must be computationally expensive and not re-
quire massive data transference between the smart-
phone and the cloud. We still intend to explore other
methods from the Android Framework that can bene-
fit from this technique.
CLOSER 2019 - 9th International Conference on Cloud Computing and Services Science
352
(a) Total Execution Time Comparison (Sao Paulo) (b) Total Execution Time Comparison (Sydney)
(c) Time to Detect Faces Comparison (Sao Paulo) (d) Time to Detect Faces Comparison (Sydney)
(e) Memory Usage Comparison (Sao Paulo) (f) Memory Usage Comparison (Sydney)
Figure 4: Comparison regarding performance and resources usage between offline processing and code offloading.
Improving the Trade-Off between Performance and Energy Saving in Mobile Devices through a Transparent Code Offloading Technique
353
REFERENCES
Akherfi, K., Gerndt, M., and Harroud, H. (2016). Mobile
cloud computing for computation offloading: Issues
and challenges. Applied computing and informatics.
Benedetto, J. I., Neyem, A., Navon, J., and Valenzuela,
G. (2017). Rethinking the mobile code offloading
paradigm: from concept to practice. In Mobile Soft-
ware Engineering and Systems (MOBILESoft), 2017
IEEE/ACM 4th International Conference on, pages
63–67. IEEE.
Chun, B.-G., Ihm, S., Maniatis, P., and Naik, M.
(2010). Clonecloud: boosting mobile device appli-
cations through cloud clone execution. arXiv preprint
arXiv:1009.3088.
de Oliveira, R. R., da Silva Schirmer, N. M., Machry,
M., and Ferreto, T. C. (2017). A transparent code
offloading technique for android devices. In Wire-
less Communications and Mobile Computing Confer-
ence (IWCMC), 2017 13th International, pages 1078–
1083. IEEE.
Flores, H., Hui, P., Tarkoma, S., Li, Y., Srirama, S., and
Buyya, R. (2015). Mobile code offloading: from con-
cept to practice and beyond. IEEE Communications
Magazine, 53(3):80–88.
Mell, P., Grance, T., et al. (2011). The nist definition of
cloud computing.
Qian, H. and Andresen, D. (2015). Extending mobile de-
vice’s battery life by offloading computation to cloud.
In Proceedings of the Second ACM International Con-
ference on Mobile Software Engineering and Systems,
pages 150–151. IEEE Press.
Sanaei, Z., Abolfazli, S., Gani, A., and Buyya, R. (2014).
Heterogeneity in mobile cloud computing: taxonomy
and open challenges. IEEE Communications Surveys
& Tutorials, 16(1):369–392.
Shiraz, M., Gani, A., Khokhar, R. H., and Buyya, R. (2013).
A review on distributed application processing frame-
works in smart mobile devices for mobile cloud com-
puting. IEEE Communications Surveys & Tutorials,
15(3):1294–1313.
Thakur, P. K. and Verma, A. (2015). Hybrid process cost
evaluation method in mobile code offloading. In
Next Generation Computing Technologies (NGCT),
2015 1st International Conference on, pages 149–152.
IEEE.
Zhou, B., Dastjerdi, A. V., Calheiros, R. N., Srirama, S. N.,
and Buyya, R. (2015). A context sensitive offload-
ing scheme for mobile cloud computing service. In
Cloud Computing (CLOUD), 2015 IEEE 8th Interna-
tional Conference on, pages 869–876. IEEE.
CLOSER 2019 - 9th International Conference on Cloud Computing and Services Science
354