Adaptive Computation Offloading in Mobile Cloud Computing

Vibha Tripathi

MIO, NYC, U.S.A.

Keywords: Mobile Cloud Computing, Computation Offloading, Data as a Service (DaaS), Platform as a Service (PaaS),

Software as a Service (SaaS), Infrastructure as a Service (IaaS), Machine Learning, Artificial Intelligence,

Augmented Reality, Internet of Things (IoT), Nash Equilibrium.

Abstract: Mobile Computing has been in use for a while now. A mobile device is a concise tool with limited

computational resources like battery, CPU and memory. Although these resources suffice the immediate

traditional needs of its user, as the mobile devices are fast turning into personal computing devices, with the

rapid development in Cloud-Based technologies like Machine Learning in the Cloud, Data as a Service,

Software as a Service, and so on there is an emergent need to implement iteratively more effective ways to

offload mobile computation to the Cloud in an on-demand, adaptable and opportunistic way. The major

issue in implementing this requirement lies in the very fact that mobile devices are location and context

sensitive, limited in battery capacity and need to be constantly reconnecting with their provider’s Base

Transceivers while still providing efficient response time to its user. In this paper, we survey this issue and a

few proposed solutions in this area and in the end; propose a model for adaptive computation offloading.

1 INTRODUCTION

Mobile devices have been around for some time

now. Mobile Computing has come of age. Cloud

Computing is rapidly growing in its own space. For

Mobile devices to be able to leverage on Cloud

Computing resources available today, there are some

constraints to consider. We cannot simply merge the

two worlds of Mobile and Cloud Computing due to

the difference in nature of a handheld or wearable

device versus a desktop machine virtually always

connected on the same Local Area Network (LAN).

Mobile devices often need to connect and

reconnect to their provider’s transceivers owing to

possible location changes.

The functional collaboration required between a

Cloud and a Mobile Network, to make a mobile

user’s Quality of Experience (QoE) seamless is

complex.

Computation offloading involves methods and

means to decide the optimal nature and amount of

computation to be delegated from a Mobile device to

the Cloud on an on-demand basis and on-the-go.

In this paper, we study the nature of some of the

complexities in Computation Offloading, proposed

solutions and their limitations and finally propose a

possible adaptive computation offloading model for

Mobile Cloud Computing.

2 CLOUD ARCHITECTURE

Cloud Computing offers ‘as a Service’ frameworks

that are used in collaborative computations. For

Mobile Computing, the service components that are

of more significance can be shown as in Fig. 1.

Figure 1: Cloud Service Layers.

The three layers are Software as a Service (SaaS)

e.g. email service, data mining and machine learning

applications, Platform as a Service (PaaS) e.g.

SaaS

PaaS

IaaS

524

Tripathi, V.

Adaptive Computation Ofﬂoading in Mobile Cloud Computing.

DOI: 10.5220/0006348505520557

In Proceedings of the 7th International Conference on Cloud Computing and Services Science (CLOSER 2017), pages 524-529

ISBN: 978-989-758-243-1

operating systems, and Infrastructure as a Service

(IaaS) e.g. storage and virtual machines.

At the core of Cloud Computing is the concept of

Hardware Virtualization – many virtual servers

participate independently with their own operating

environment but with the same physical layer.

Mobile Networks are laid out architecturally to

provide an illusion of pervasive uninterrupted

network connection to the end user while making

location and reconnections appear seamless most of

the times. The architecture is described in the

Section 2.1.

2.1 Mobile Cloud Computing and

Offloading Computations

Figure 2: The figure shows how handheld devices switch

connections to their base transceivers based on their

location in a cell.

Computation and Data Intensive Apps: Some mobile

applications may need to perform computation and

data-intensive tasks like voice and speech analysis,

facial expression analysis (Google Cloud Vision

API), Augmented Reality etc.

Cloud based mobile apps can allow the end users to

leverage the vast resources of the Cloud and run in a

virtual environment, by offloading the data-intensive

computation components to the Cloud and provide a

better response time from the Cloud as opposed to

local device-based computations when

communication is light (Jesus Zambrano, 2015).

Mobile Battery Life: With the rapid increase in the

multimedia content and computation intensive

games being made available for mobile system, one

of the primary constraints while considering

computation offloading is the mobile battery life.

Battery technology has until now been unable to

catch up with the fast growing battery consumption

on mobile devices.

(Karthik Kumar et al, 2012) propose a formula to

dynamically decide on computation offloading of

certain energy-intensive computations with a view to

benefit from offloading as opposed to local mobile

based computation which can drain out the entire

battery.

Along with the computational components, however,

we also need to consider the location and context

awareness of these mobile devices.

Location and Context Awareness of Mobile Devices:

Mobile devices have sensors that can collect

information about the location of the device, some

context metrics about the users (e.g. Fitbit) for

example to recommend services, customize answers

to user queries, advertise local businesses, provide

individualized search results etc. (Gabriel Orsini et

al, 2015) discuss the requirements, design and a pre-

phasing or partitioning for computational offloading

to arrive at an optimal context aware solution.

A typical Cloud-enabled Mobile device would have

the below minimum requirements:

• An uninterrupted connection to a Cloud

service provider

• An efficient computation offloading model

that takes context awareness as well as

battery usage efficiency into account

• An effective computation component

distribution model

2.2 Offloading Model

Fig. 3 shows how the current offloading process

appears considering the major participants. However

the components involved in a typical mobile

application computation offloading to Cloud can be

categorized as below:

• Surrogate – a computing node or a virtual

environment made available by the Cloud

service provider where the offloaded code

can run

• Partitioner – divides and sorts out the

application components to be offloaded for

Cloud-based computing, this could be static

or dynamic (Porras, J. Riva, 2009)

• Context Monitor – Monitors and provides

context related information about the

device, the available surrogates in the local

area, battery status, connectivity etc.

• Solver – uses the information from Context

Monitor to decide which surrogate to

offload the computation the Partitioner

decides to offload

Mobile

Switching

Service

Base

Transceiver

Adaptive Computation Ofﬂoading in Mobile Cloud Computing

525

• Coordinator – authenticates and securely

sends computation over to surrogate,

receives and presents the results from the

surrogate

….. Cloud

Surrogates

Figure 3: Offloading Mobile Computation.

2.2.1 Offloading Granularity and Surrogate

Selection

Several approaches are in use at the moment, with

each having its own pros and cons. Traditional

remote method invocation (RMI) and client-server

model for offloaded computation keep the object-

oriented-ness for mobile app designers but require

synchronization of variables which may mean error-

prone context-insensitive computation results.

Virtual Machine based offloading can be a language

based VM for the mobile app or a mirrored mobile

computing environment with the ability to run any

app in a given mobile environment.

(Xu, Chen et al, 2015)

propose a game theoretic

approach to decision making problem in

computation offloading amongst multiple mobile

device users. They formulate the problem as a multi-

user computation offloading game and prove that it

always admits Nash equilibrium. They also propose

a distributed offloading algorithm that can achieve

this equilibrium. However, the dynamicity of mobile

users frequently joining and leaving the game on-

the-fly has yet to be considered in this approach.

None of the above solutions however, make it easy

for the Partitioner to decide the granularity of

offloading computation components on-the-fly or for

the Coordinator to decide which Surrogate to offload

to.

2.2.2 Opportunistic Partitioning

For an opportunistic Partitioner the application

should be able to compute the cost of local versus

cloud-based computation based on the context

information e.g. battery usage, available bandwidth

and quality of network connection. Derived from the

report by (S. Pachamuthu & Kumar, 2016) a simple

formula to compute the cost can be as below:

α (Tcloud - Tlocal )/Tlocal + (1-α )(Ecloud - Elocal)/Elocal

Where α denotes a weight between 0 and 1 towards

computing time T or energy consumed E being

assigned that weight.

Mobile Apps that can run optimizations to decide

what parts of their computation should be offloaded

and what should be run locally can either come with

pre-defined static partitioning of the program into

components while some others can run this

optimization itself in the cloud.

2.2.3 Dynamic Surrogate Selection

In Section 2.2.1, we mentioned the complexity with

Coordinator to decide which Surrogate to connect to

for offloading. Considering the mobility of a

handheld, we acknowledge another level of

complexity to the decision making – the mobile

device should potentially be able to connect and

offload to multiple Surrogates and able to receive

and reconstruct results from them dynamically.

2.3 Current Solutions

2.3.1 Native Mobile Cloud Computing

Some solutions in use today have been designed

specifically to solve the offloading from mobile

devices. These solutions fall in two broad categories.

The first category is, where the entire mobile

application or parts of it are run in Virtual Machines

mimicking the mobile device on servers to decide

which threads of methods should be offloaded when

running from a mobile device. CloneCloud (Chun,

B.G, 2011) and MAUI (Cuervo, E., 2010) are such

examples. The application running on the server’s

VM is profiled for its cost of computation.

In the second category, the computation

components are analysed a-priori and decisions are

made to offload more computation intensive tasks to

Surrogates. This solution doesn’t take into account

the context sensitivity of the mobile device and

hence results can be unsatisfactory.

1) Offload

2) Request computation

3) Receive results

Offloaded Computation

Com

onent Execution

CLOSER 2017 - 7th International Conference on Cloud Computing and Services Science

526

Surrogate2

Surrogate3

Surrogate1

Hub

(DaaS)

2.3.2 Frameworks and Domain Specific

Languages

There are many programming languages in market

today, specifically designed to solve problems in

domains like mobile commerce etc. These languages

are often required to run in their specialized run time

environment. While this solution follows an agile

approach as the developers are focussed in the

domain, it does require developers to niche their

expertise for a specific language and run time

environment.

Various frameworks and platforms like Java

Remote Method Invocation provide proven

methodologies for distributed execution of

applications, thereby letting the developer focus on

the application functionalities. However, these

conventional standard frameworks are not flexible

enough to support mobile computation offloading in

rapidly varying contexts.

2.3.3 Distributed Virtual Machines

In this approach multiple VMs run on multiple

nodes in a network and run computation components

with a common global state of the application

(Coulouris, G., Dollimore et al, 2012). However this

solution requires the mobile device to support the

distribution of the tasks. There are proposals for

using embedded devices in routinely used objects

like shoes, watches and so on for computation

offloading on personal distributed systems

(Niroshinie Fernando. et al, 2013)

(Verbelen, T., 2012) propose a new architectural

paradigm by converging mobile computing, Internet

of Things (IoT) and Cloud Computing. A Cloudlet

sits between the Cloud and the Mobile or IoT device

and can serve as a small data centre bringing the

cloud closer to the device by providing a PaaS layer

which can support wearable cognitive assistance,

making possible new flurry of applications based on

Artificial Intelligence and Augmented Reality for

the Mobile or IoT user.

2.3.4 Other Solutions

Some solutions take the application partitioning and

component execution to the system level. Mobile

applications in this approach are designed to assure

that computationally intensive parts of the app can

be run in distributed environment. These solutions

do not however do part by part offloading from the

mobile device and instead consider the application in

entirety as per the design.

3 TOWARD AN ADAPTIVE

OFFLOADING MODEL

We propose that an effective cloud-enabled Mobile

App should be elastic in its decision-making of what

and where to offload its computation.

Additionally, the Mobile Cloud Computing

infrastructure should be able to provide a dynamic

selection of Surrogates that can pass on partial or

completed results from offloaded computation based

on the changing context and location of the device.

The benefit in our model arises from the Surrogates

providing a virtual PaaS by partially or completely

participating in the computation of offloaded

requests. The Mobile device may not be aware of

this virtual network of Surrogates and may receive

the results of its offloaded computation from any

one of the participating Surrogates.

3.1 Virtual Network of Surrogates

Figure 4: Virtual Networks of Surrogates.

We propose a model for surrogates that connect

amongst themselves via a cloud-based hub that

provides Data as a Service. When a mobile device

selects for example Surrogate1 in the Fig.4, based on

the application’s requirements, it either precomputes

partitioning or offloads the execution components

entirely to that Surrogate, any computation results

and context information is sent to the DaaS Hub,

which upon request provides either partially

computed or completed results to another Surrogate

which the mobile device may have selected upon

location change.

Context information and the cost of offloading may

also have changed with the location, and hence that

should be taken into consideration when selecting

the next Surrogate.

Adaptive Computation Ofﬂoading in Mobile Cloud Computing

527

Having a Hub providing DaaS, resolves several

issues like availability, since some computation

results may always be available for the mobile

device via the Hub once at least the first Surrogate

was selected for offloading. It also provides data

security as most of the data intensive transactions

would be limited within the Surrogate network.

The above proposed model also eases

maintainability of mobile applications as developers

can focus on the application itself while most of the

intensive computation, including the optimization

for Partitioner to decide what components to offload,

can be run in selected Surrogate once and all that is

exchanged with subsequently selected Surrogates

could be Context Sensitive information.

3.2 Multi-Surrogate Distributed

Offload

The model proposed in Section 3.1 would also allow

for more scalability as multiple Surrogates can be

employed in an elastic manner to compute a

distributed task intensive application. The selected

Surrogates can, not only compete for their

candidacies based on the Virtual Environment they

can provide to run parts of computation from a

mobile app, they can also utilize results from

participating Surrogates.

Figure 5: Distributed Computation Offloading.

This way we can apply heterogeneously specialized

VMs, for example to provide Cloud Based Machine

Learning routines, Video Games and so on.

3.3 Centralized Offload and

Distribution

Another approach with the abovementioned virtual

network of surrogates can be as shown in Fig. 6 for

the mobile application to offload centrally to the first

selected Surrogate which acts as a master and

decides to distribute tasks to other Surrogates based

on the resource availability. Big Data solutions like

Hadoop internally employ such distribution of tasks.

Figure 6: Centralized Computation Offloading.

4 CONCLUSION

The main challenges in the field of Mobile Cloud

Computing include Context Sensitive Computation

Offloading. In order for the Mobile users to leverage

from the vast resources and functionalities offered

by rapidly advancing Cloud Computing

technologies, much further research is required in

the area of computation offloading.

We studied state of the art Mobile Cloud

Computing, current challenges and a few proposed

solutions. Based on the study we proposed a model

for opportunistic and adaptive offloading with a

focus on the context awareness of computation

intensive mobile applications. Further work is

required to run evaluations with computation

intensive mobile applications using the model and

fine tune the model.

In conclusion, Mobile Cloud Computing is still in its

nascent state and while much of the traditional

distributed application approaches can be utilized to

some extent in the context of MCC, certain areas

like adaptive context sensitive computation

offloading require further research.

REFERENCES

Porras, J. Riva, O., Kristensen, M.D. Dynamic resource

management and cyber foraging. In: Middleware for

Network Eccentric and Mobile Applications. Springer;

2009, p. 349–368.

Computation

Hub

(DaaS)

Computation

Offload

Hub

(DaaS)

Computation

Distribution

CLOSER 2017 - 7th International Conference on Cloud Computing and Services Science

528

Coulouris, G., Dollimore, J., Kindberg, T.. Distributed

Systems: Concepts and Design. Boston, USA:

Addison-Wesley; 5 ed.; 2012.

Verbelen, T., Simoens, P., De Turck, F., Dhoedt, B..

Cloudlets: Bringing the cloud to the mobile user. In:

Proceedings of the Third ACM Workshop on Mobile

Cloud Computing and Services; MCS ’12. New York,

NY, USA: ACM. ISBN 978-1-4503-1319-3; 2012, p.

29–36.

Gabriel Orsini, Dirk Bade, Winfried Lamersdorf Context-

Aware Computation Offloading for Mobile Cloud

Computing: Requirements Analysis, Survey and

Design Guideline, MobiSPC 2015

Karthik Kumar, Jibang Liu, Yung-Hsiang Lu ,

Bharat Bhargava: A Survey of Computation Offloading

for Mobile Systems, Mobile Netw Appl DOI

10.1007/s11036-012-0368-0, 2012

Jesus Zambrano, Mobile Cloud Computing: Offloading

Mobile Processing to the Cloud, University of North

Florida 2015

Niroshinie Fernando, Seng W. Loke, Wenny Rahayu;

Mobile cloud computing: A survey, Elsevier: Future

Generation Computer Systems Volume 29, Issue 1,

January 2013, Pages 84–106

Chun, B.G., Ihm, S., Maniatis, P., Naik, M., Patti, A..

CloneCloud: elastic execution between mobile device

and cloud. In: Proceedings of the 6. European

Conference on Computer Systems. 2011, p. 301–314.

Cuervo, E., Balasubramanian, A., Cho, D., Wolman, A.,

Saroiu, S., Chandra, R., et al. MAUI: Making

smartphones last longer with code offload. In: ACM

MobiSys 2010

Xu Chen, Lei Jiao, Wenzhong Li, and Xiaoming Fu

Efficient Multi-User Computation Offloading for

Mobile-Edge Cloud Computing, IEEE/ACM

Transactions on Networking 2015

S. Pachamuthu & Kumar Rochester Institute of

Technology, Rochester, New York Cost Evaluation of

Computation Offloading on Mobile Devices, 2016

Adaptive Computation Ofﬂoading in Mobile Cloud Computing

529