1 Introduction

Smart mobiles have been becoming popular because of their increasing processing capabilities, convenience, and portability. It is predictable that mobile devices will replace the traditional laptop and desktop computers in the near future [1]. However, mobile systems have inherent limitations, including short battery life and limited computing and storage capabilities [2]. Enhancing capabilities of mobile devices can be achieved by a new computing paradigm called mobile cloud computing (MCC) [3].

MCC enhances and optimizes mobile devices’ computing capability and it can alleviate storage and mobility problems of mobile devices. It brings new types of services and facilities for mobile users to take full advantages of cloud computing. A number of MCC solutions have been proposed, such as Misco [4], Clonecloud [5], and Hyrax [6]. March [7] developed a component based mobile application framework leveraging cloud resources. Lu [8] proposed a conceptual architecture to perform screen rendering tasks inside the cloud. SOPHRA [9] uses a cloud centric middleware to enable mobiles to host web services in the mHealth domain. In [1], the author presented a new mobile cloud architecture, leveraging the collaboration between mobile devices and conventional desktop or laptop computers to empower mobile computing, particularly aimed at optimizing data distribution and resources’ utilizations so that expected Quality-of-Service (QoS) requirements can be met. They also proposed an algorithm to select an optimal resource allocation strategy to satisfy various service level agreements (SLA). MCC has been used in various applications including mobile commerce, mobile learning, and mobile gaming [3].

Enhancing capabilities of small devices in MCC can be achieved by migrating applications, and offloading computing intensive or data intensive tasks to proximate or remote cloud nodes [10, 11], which is called task migration in MCC. In some occasions, the most effective application migrations should be able to simultaneously meet multiple user requirements [12]. A computation migration is also confronted with several key challenges such as context awareness and remote execution control [13]. Considering all of the problems, we need a good migration scheme about which part to migrate, when and where to migrate. So decision making for task migration is an important issue which can affect application executions including security, performance, power consumption of mobile devices [3] and of whole systems, and hence affect experiences of mobile users.

In this paper, we present a relatively comprehensive survey on the decision making of task migrations. The motivation is to discuss some important issues in task migration processes including what kind of decision factors should be considered and what kind of decision mechanisms are used. Besides, we present some challenges in order to point out future research directions.

The rest of the paper is structured as follows. Section 2 gives an overview of mobile cloud computing and introduces the task migration in MCC including generally steps and architectures for task migrations. A survey of related work and decision-making approaches for task migration is given in Sect. 3. In Sect. 4, A comprehensive comparison and discussions to migration decision approaches are presented. Some challenges of the task migration in MCC are listed in Sect. 5. Section 6 concludes the paper.

2 Overview of task migration in MCC

The MCC forum defines MCC as follows:

‘Mobile cloud computing at its simplest refers to an infrastructure where both the data storage and data processing happen outside of the mobile device. Mobile cloud applications move the computing power and data storage away from mobile phones and into the cloud, bringing applications and MC to not just smart phone users but a much broader range of mobile subscribers’ [14].

There are some other definitions, such as Aepona [15] and Liu [16]. We can describe the general idea of mobile cloud computing as the following architecture shown in Fig. 1. Potentially, MCC has a number of advantages such as extending battery lifetime, improving data storage capacity, enhancing performance, and dynamic provisioning [3].

Fig. 1
figure 1

General idea of mobile cloud computing

Full size image

2.1 Reasons and steps to do task migrations

There are a number of reasons to conduct task migrations. First, mobile applications are getting more complicated and more computing intensive [17]. Second, mobile users would like to obtain better performance when using mobile devices which have inherent limitations with respect to a built-in memory, battery, storage capability, and furthermore, poor bandwidth [2]. In a word, the reason for doing task migrations in mobile cloud environment is that there are mismatches between requirements and the mobile devices’ capabilities.

In order to make migrations meaningful, some factors need to be considered. We classify these factors into four categories: applications’ characteristics, mobile devices’ characteristics, environment factors, and user’s preferences and requirements. These factors can help to decide whether a migration should be conducted and how to proceed task migrations.

Generally, the steps of task migrations are as follows:

  1. 1.

    Looking for available resources and services which mobile devices can offload computing tasks to [18, 19].

  2. 2.

    Monitoring environment context information such as status of surrounding networks and relevant devices, predicting what are needed through some sort of reasoning approaches [20–23].

  3. 3.

    Conducting application partitions at a suitable granularity [24] and making decision for tasks scheduling with appropriate algorithms [25, 26].

  4. 4.

    Remote execution control for applications needs to be conducted. This includes how to interact with mobile clients, such as getting users’ input and returning results to mobile ends.

The first and second steps make preparation for task migration which are used to collect useful information. The third step is the core for task migrations, namely decision making of task migrations. Effective decisions are important. The last one makes sure that applications can be correctly executed even after the migrations.

2.2 Task migration architectures of MCC

A task migration architecture means what type of resources mobile devices use to enhance their abilities and the way in which mobile devices are connected to and interact with cloud resources. According to the resources organization, existing task migration architectures can be classified as the following four types:

  1. 1.

    Public cloud: Mobile devices offload tasks to public resources comprised of stationary servers located far from the mobile nodes or proximate centralized desktops and laptops accessible via the Internet. The migrated tasks or portions of applications can be executed remotely in the cloud. Some task migration solutions use this architecture to enhance capabilities of mobile devices and get better execution of applications. In MAUI [27], mobile devices offload code to a infrastructure composed of immobile MAUI servers. Smart phones use remote procedure call to communicate with MAUI servers. Clonecloud [5] also uses a computational public cloud architecture to transfer mobile applications. It transforms a single machine execution (e.g., smart phone computation) into a distributed execution (e.g., smart phone plus cloud computation) in which the resource intensive parts of the execution are run in powerful clones in the cloud. Public cloud is highly available, scalable, and resource elastic with powerful computing capabilities while mobile phones are resource constrained. A novel MCC architecture was proposed in [1]. It leverages the collaboration between mobile devices and conventional desktop or laptop computers to empower mobile computing to optimize data distribution and utilization of CC resources. As said in [17], the disparity in capabilities between public cloud and smart phones is high and persistent. Therefore, leveraging public cloud to augment mobile devices can benefit mobile application executions pretty much. But the communication time and cost must be considered for migrations because accesses to the cloud may incur a long latency due to wide area network (WAN) delays.

  2. 2.

    Cloudlet: A Cloudlet [28] is one computer or cluster of computers that is well connected to the Internet and available to nearby mobile devices. It is resource rich and proximate. Due to the proximity, the end to end response time of applications executing within it is little (a few milliseconds) and predictable. Cloudlet simplifies the challenge of meeting peak bandwidth demands and gives users better experience especially for realtime mobile applications. But using Cloudlets also faces to some practical issues such as reliability problems. Cloud-Vision [29] proposed a mobile-Cloudlet-cloud architecture named MOCHA to do realtime face recognition. The Cloudlet determines how to partition computation among itself and multiple servers in the cloud to optimize the overall Quality of Service (QoS). The implementation and result demonstrate that Cloudlet is technically feasible and high powered to provide benefits to mobile device face recognition applications considering its higher computational power with minimal latencies .

  3. 3.

    Cloud of Mobile Network Operator (MNO): A mobile network operator or MNO is a provider of wireless communications that owns or controls all of the elements necessary to sell and deliver services to an end user including radio spectrum allocation, wireless network infrastructures, and provisioning of computer systems and marketing [30]. There are many MNOs like China Mobile and AT&T which has provided cloud services [31]. As said in [32], a MNO can be a cloud provider by deploying cloud concepts including SaaS, IaaS. A MNO can be a cloud broker which plays as a service mediator between cloud providers and mobile users, and from the other aspect, a broker across other MNOs. Cloud of MNOs is widely covered, safe, and reliable. Mirror [33] keeps a mirror for each smart phone on a computing infrastructure within 3G networks. Paper [32] proposed a service based arbitrated multi-tier infrastructure for mobile cloud computing. MNOs are used as arbitrators between front end (mobile users) and back end (cloud service providers) to do resource and service allocation for more efficient executions. A MNO and its authorized dealers compose the infrastructure together with clouds.

  4. 4.

    Cloud of mobile devices: In the former three architectures, mobile devices act as thin clients which utilize external resources in cloud computing platforms to process computing intensive or data intensive tasks. But we cannot avoid situations that an access to these platforms may be not available and/or it is too expensive to access them. Cloud of mobile devices is a solution to these situations. Mobile devices play the role of resource provider, those which are in vicinity of users and in a stable mode form a virtual cloud for task offloadings. In Hyrax [6], the author introduced a concept of using mobile devices as resource providers. Similar work has been done in grid computing. Black and Edgar [34] demonstrated the feasibility of using mobile devices as part of a grid. With cloud of mobile devices, users can share resources and execute jobs among the devices to gain better experience. This is very suitable for some scenarios like group tourism. [35] created a cloud among devices in vicinity taking advantage of the pervasiveness of mobile devices.

Table 1 shows a simple comparison of the four architectures. Each architecture has its specific characteristics and can meet different task migration requirements. When doing task migrations, we can choose a suitable mobile cloud architecture according to migration scenario and objectives.

Table 1 Task migration architectures of MCC
Full size table

Apparently MCC has been a hot topic to augment mobile devices capabilities. There have been many research efforts on mobile cloud environment, as discussed in [3, 24] and many solutions for task migrations in mobile cloud computing as represented in the following Sect. 3.

3 Decision making for task migration

Section 2 delivers an overview for task migration in MCC including the reasons to do task migrations and architectures existing in MCC. To ensure efficient program executions, a task migration needs an intelligent decision making way to adapt to dynamic changes in mobile cloud computing environments and achieve better execution performance. In this section, we focus on decision making for task migration in MCC and some specific existing approaches of task migrations will be talked about with some detailed descriptions.

There have been many solutions about decision making for task migration in mobile cloud computing. They are all for augmenting mobile devices and making applications better executed, but they are different from each other in decision objectives and decision algorithms. A significant difference is that a migration decision can be static or dynamic. Now in the following is some approaches with the two different decision making ways. We discuss each approach with a detailed description.

3.1 Static task migration decision approaches

  1. 1.

    Wishbone: Wishbone [36] is an application partition system based on a data flow graph of operators. Its partitioning algorithm models the application partition problem as an integer linear program (ILP). It uses an ILP solver to solve the optimal graph partitioning problem to get a partition scheme and then assigns tasks to a central server or a sensor node. In this way it can reduce mobile devices’ cost as much as possible, which is a linear combination of network communication overhead and CPU load. In Wishbone, the partition approach is only suitable to those applications which can be expressed as a data flow graph.

  2. 2.

    VM placement/migration: Network I/O performance would affect application performance significantly. Therefore, [37] proposes a network aware virtual machine placement and migration approach in cloud computing environment. It aims to minimize data transfer time of an application. In the proposed VM migration policy shown in Fig. 2, a VM migration is triggered when the data transfer time crosses a certain threshold due to unstable networks. Then, the next optimized physical machine which has minimal access time is chosen according to the current network condition, and the VM is migrated to this physical machine for better performance. Experiment result shows that average task completion time is decreased with task migrations. A precondition of the proposed approach is that host storage and network bandwidth must be known in advance.

  3. 3.

    Clonecloud: Clonecloud [5] is a system dynamically migrating mobile applications to cloud. It uses remote cloud resources and partitions applications at a thread level to obtain a fine granularity and flexibility. A mathematical optimizer chooses migration points that optimize the objective (such as total execution time or mobile device’s energy consumption) and makes dynamic profiling to build a cost model. The runtime system chooses which partition to use. After the partition and migration, the application is under distributed execution (parts of the execution are seamlessly offloaded to computational cloud, while rest are still executed locally), and the result will be back to the mobile device to be integrated.

  4. 4.

    Task-resource Scheduling: In [38], the author gave a solution to task-resource scheduling problem in cloud computing system. It describes an application work flow as a directed acyclic graph comprised with tasks and their data dependency relationships. The task-resources scheduling problem is to find a scheduling scheme to minimize total execution cost of applications running on the resources provided by cloud service providers. Total execution cost is the sum of computation time and file transmission time. This problem is reduced to a satisfiability problem and then solved with a SAT solver. The overall process is as shown in Fig. 3.

  5. 5.

    Task Scheduling with DVFS: A novel algorithm for the MCC task scheduling problem is presented in [39]. The algorithm is designed to minimize total energy consumption of an application in a mobile device under a hard constraint on the application completion time. An application is represented by a directed acyclic task graph \(G =(V; E)\). Each node \(v_i\in V\) represents a task, and a directed edge \(e(v_i;v_j)\in E\) represents the precedence constraint. And the task-precedence requirements are considered and customized. There are three steps when doing task scheduling. In the first step, a minimal delay task scheduling is generated without considering energy consumption of the mobile device. The initial scheduling algorithm takes into account a joint scheduling of tasks on local cores, wireless communication channels, and cloud nodes. And then the task migration algorithm performs energy reduction in the second step by migrating tasks to the cloud or other local cores that can bring great energy reduction without violating the application completion time constraint. In the third step, dynamic voltage and frequency scaling (DVFS) technique is used to further reduce energy consumption by making a suitable choice of the operating frequency for each local core. To avoid high time complexity, the author proposed a linear time rescheduling algorithm for the task migration. When it comes to the overall computation complexity of the MCC task scheduling algorithm, it is \(O(N ^3 \times K)\) where N is the number of tasks and K is the number of cores.

    From above descriptions about these static decision making approaches, it is not hard to find that static migration decisions have the advantages of simplicity and low overhead during executions. Compared with them, dynamic decision approaches make more consideration to different runtime conditions and are more adopted by current researches.

Fig. 2
figure 2

Network aware VM migration approach

Full size image
Fig. 3
figure 3

Using a SAT solver to do task-resource scheduling

Full size image

3.2 Dynamic task migration decision approaches

  1. 6.

    Parametric Analysis: The work of [40] proposes a parametric analysis approach to achieve an optimal program partitioning and scheduling for computation offloading. Applications are divided into tasks according to task control flow graphs (TCFG). And then based on the optimal program partitioning that corresponds to the current values of runtime parameters, the program is transformed into a distributed program which at runtime self-schedules its computation tasks to either a mobile device or a server. The overall process is as shown in Fig. 4. There is a trade-off between computation workload and communication cost when making a partition and migration decision. It obtains program computation workload and communication cost with cost analysis and expresses them as functions of runtime parameters, and then the parametric partitioning algorithm will find the optimal program partitioning corresponding to different ranges of runtime parameters.

  2. 7.

    (k + 1) Multi-Constraint Partitioning Algorithm: The paper [41] proposes an adaptive (k + 1) partition algorithm for application offloading. Statically, it partitions an application using a graph cutting algorithm at class level according to the application’s features (a dynamic multi-cost graph). According to the partition scheme, the multi-constraint partitioning algorithm obtains a best distribution of the application where one un-offloadable partition will run locally on the mobile device and k offloadable partitions are distributed to k surrogates.

  3. 8.

    MAUI: [27] is an architecture where mobile program code can be offloaded to a powerful infrastructure. MAUI decides at runtime which methods should be remotely executed, driven by an optimization engine that achieves the best energy savings possibility under the mobile device’s current connectivity constrains. To offload an application, it uses collected samples of CPU utilization and the corresponding smart phone energy consumption to build a simple linear model, using the least-square linear regression. Then, it can predict the energy consumption of task executions and optimize the current model using machine learning. Based on the energy consumption model, decision making of application partition and migration is dynamically performed at the method level. High level overview of the MAUI architecture mainly includes two parts, smart phone and MAUI server. Each end has a proxy, a profiler, and a solver. The profilers are used to do device and program profilings, and the MAUI solver uses data collected by the MAUI profiler as input to determine which remotable methods should be executed locally and which should be executed remotely.

  4. 9.

    Darwin: In [42], the authors introduced an integrated automation framework called Darwin. It enables workload migrations in a source-target mode. The objectives are to enhance the speed, reduce the cost and inherent risk of performing migrations. During a workload migration, it firstly discovers the source environment, makes analysis to select migration target candidates, and then creates a workload migration map. Darwin configures the target platform(s) environment and application components to carry out the migration at last. As a framework, Darwin is able to architecturally support different variations but still has gaps such as lack of open virtualization format (OVF) integration with the request for migration XML specification.

  5. 10.

    EOA: [43] proposes a novel efficient online algorithm (EOA) to solve the migration problem of batch jobs among data centers (DCs). EOA can handle future variabilities and uncertainties of energy sources, and can make fundamental trade-off between energy and bandwidth costs of migrating application data and states. Online job migrations are defined as minimum value problems of the total operation cost (sums of the energy costs at a DC and the bandwidth costs of migrating jobs among DCs) in the cloud. As a computationally efficient algorithm, the complexity of EOA algorithm is O(logn) where n is the total number of servers in the cloud.

  6. 11.

    Cloud-Vision: The work [29] introduces an application migration approach to increasing mobiles’ performance (response time) by using proximate Cloudlet. A Cloudlet receives processing intensive tasks from mobile devices and then assigns them to itself or remote cloud nodes. It has two schemes for task migrations: equally distribution and greedy style. Apparently in the equally distribution scheme tasks are equally distributed among available cloud servers (or the Cloudlet). In the greedy style scheme, a migration decision is made according to the time of finishing the tasks. The latter approach has better performance, especially when the communication delays of servers are different. But a weak point is that it is difficult to know the communication delays because of the environment’s dynamics.

  7. 12.

    LALTM: The author of [44] proposed a Java byte code transformation technique for realizing task migrations which aims at minimizing the overhead. It presents how to do migrations for one application focusing on state capturing before a migration and state restoring after execution results return. A migration is triggered by application itself when the execution has reached some special states such as exceptions. The migration mechanism makes use of asynchronous exceptions and byte code instrumentation to perform state capturing which can avoid to poll the migration state and check the migration state repeatedly after every function returns. In the state restoring, a technique, namely Twin Method Hierarchy, is used to minimize the overhead induced by code restorations. The mobile device transfers just enough portion of the stack, methods, and objects to the cloud node. This minimizes the overhead of migrating tasks and allows more flexible migration paths.

  8. 13.

    EMSO: [45] presents an energy efficient multisite offloading (EMSO) algorithm that saves energy and reduces time consumption. EMSO divides an application by depth-first search based on a weight object relation graph (WORG). After static analysis to the application, partition and offloading are dynamically performed and can adapt to environment changes. The test result of the system shows that EMSO can significantly reduce energy and time consumption.

  9. 14.

    MDP migration: In [46], the authors modeled the service migration problem for Follow Me Cloud (FMC) using a Markov decision process (MDP) and formulated a decision policy to determine whether and where to do service migrations according to the trade-off between migration cost and user perceived quality. The overall process of Markov decision process is as shown in Fig. 5. FMC is a network architecture which enables service mobility across federated data centers (DCs). Following the mobility of a mobile user, the service located in a given DC is migrated at every time when an optimal DC is available. In the MDP model, given a state (current service location), there are a set of actions (migrate or not) which are associated with transitions to other states. For each transition, there is a corresponding reward and cost function. The decision policy is used to find an optimal migration policy which has the max discounted total reward. Numerical results show that the proposed service migration decision mechanism always achieves the maximum expected reward.

  10. 15.

    OSGi-PC: OSGi-PC proposed in [47] is based on the de facto Java component standard called OSGi. In OSGi-PC, the work for partitioning application is not yet considered but as it uses a component based approach, the support for application partitioning will not be hard to add. The offloading is done via component migrations across any kind of cloud nodes including the migration from powerful nodes to mobile client nodes.

  11. 16.

    Mobiles on Cloud Nine: The paper [48] proposes a solution for task migration in cloud computing systems. Some online migration strategies are developed to make migration decisions according to instantaneous load of the system and estimated execution time of tasks. Following cloud scenarios, migration policies vary from task centric migration where migration decisions are made by the task itself for minimizing its own execution time to cloud wide task migration decided by the cloud provider considering the whole system’s performance. When making task migration decisions, three factors are explicitly taken into account: the multi-tenancy effect resulting from interaction of co-located VMs in a server, the cost of migrating source code and input data of the task, and the cost of transferring final result to the user. A migration should be occurred only if it is beneficial for the processing time of a task.

  12. 17.

    Jade: Jade [49] is an energy aware computation offloading system for Android mobile devices. It transports computation which contains remotable tasks from an Android device to servers running on the cloud. At runtime, Jade gathers devices’ information and tasks’ information and uses multi-level scheduling algorithm to offload tasks and balance the workload of servers. Jade was evaluated with two applications (face detection and path finding), and the result showed that it can effectively reduce up to 35 % of average power consumption for mobile devices.

  13. 18.

    DCOG: Decentralized computation offloading game (DCOG) [50] is a computing offloading decision mechanism. The authors considered communication and computation aspects of users’ tasks and formulated the computation offloading decision-making problem among multiple mobile device users as a decentralized computation offloading game which admits a Nash equilibrium. Within the offloading decision mechanism, the authors used a slotted time structure to do interference measurement and to solve the decision update contention problem cyclically and finally determined whether users’ tasks should be executed locally or migrated to the cloud.

  14. 19.

    DECM: [51] proposes a dynamic energy aware Cloudlet based mobile cloud computing model (DECM) which uses Cloudlets with dynamic programming to reduce additional energy consumption during wireless communication. DECM manages and optimizes the cloud based infrastructure usages and services. A dynamic Cloudlet is the main component of DECM to find better connections (connecting to another Cloudlet or cloud or itself) for users. DECM uses recursion to solve service migration problems.

Fig. 4
figure 4

Parametric analysis for adaptive computation offloading

Full size image
Fig. 5
figure 5

Markov decision process (MDP) for task migration in Follow Me Cloud

Full size image

4 Comparisons of decision-making approaches

In Table 2, we summarize the decision-making approaches for task migrations aforementioned from the following aspects: the scenario of task migration problem, the algorithm proposed to make a decision for task migration, the migration granularity, whether the decision making is dynamic or static, what factors are considered when making decisions for task migrations, the complexity of the decision algorithm, and finally the decision mode in which centralized or distributed decision making is adopted.

Table 2 Comparisons of decision-making approaches in MCC
Full size table

4.1 Decision-making algorithms for task migration in MCC

There are a number of application migration algorithms proposed so far. A common approach is graph based partitioning such as HELVM [41] and operation data flow graph partitioning in [36] which can make better analyses to an application and get a migration scheme easily. Some other partition methods can be used under certain circumstances such as a SAT solver which is used to solve SAT problems transformed from minimization problems [38]. MDP [46] is associated with state transitions of a task migration, and DVFS algorithm [39] is used to reduce energy consumption as described in Table 2. Each decision-making algorithm has complexity different from others which leads to different migration overheads [44].

4.2 Static or dynamic task migration decisions

Some of the algorithms make migration decision statically before an application’s execution [36, 39]. Therefore, a program is partitioned during development, and task migration is usually performed only once within static decision approaches. Some migration decisions are made in a dynamic way taking into account the runtime environment and application executions. Task migrations usually are triggered when program executions have reached some special states such as in [37] or an attribute value exceeds a threshold such as in [46].

Static algorithms are relatively simple and lead to low overhead during executions without monitoring computing contexts, but there are limitations considering dynamic situations, such as static decision approaches are valid only when parameters can be accurately predicted in advance. In contrast, dynamic decisions can adapt to different runtime conditions and can be particularly useful since most information of computing environment cannot be known in advance due to the dynamic nature of mobile cloud computing environments (fluctuating network bandwidth, constantly changing memory and CPU). However, dynamic migration decisions may lead to higher runtime overhead.

4.3 Decision mode for task migrations

Task migration mechanisms may be centralized whose decisions are made by a central migration manager or processor considering the whole system’s condition or performance [46, 47]. These centralized decision-making approaches are relatively simple but may cause bad performance to one specific computing node or task. Currently, there are some distributed decision approaches proposed such as [44, 48] in which the decision making responsibility belongs to several nodes for better resources utilization. Servers or tasks possibly make autonomous decisions independently or make cooperative decisions, not only considering its own objectives but also that of other nodes.

4.4 Decision factors considered for meaningful migrations

In order to make migrations meaningful, some factors need to be considered when making decisions as shown in Fig. 6.

  1. 1.

    Applications characteristics. Only computing intensive or communication intensive applications and applications requiring a large amount of data which is remotely located need to do task migrations. Otherwise, the benefit is not enough to offset the cost of migration produces by communication or latency. Therefore, when deciding whether and where to do migration, we must consider some applications’ or tasks’ information such as data volume, and computing time [38, 41, 45, 50].

  2. 2.

    Mobile devices’ and cloud’s characteristics. Mobile cloud environment is heterogeneous (in terms of the hardwares, platforms, the services, and networks). We need to know status and features of nodes, such as performance, availabilities, reliabilities, distance, and elasticity and costs. A migration should be performed when there is no enough resources to support applications’ execution in mobile or cloud nodes, or in order to have better execution performance. These characteristics include CPU processing capacities of both mobile devices and cloud [36, 42, 49], mobile device energy information and mobility [39, 46], and load and resource utilization such as memory and storage on cloud [43, 47, 48].

  3. 3.

    Network environment. Mobile cloud computing is developed depending on the utilization of networks. There is no doubt that communication technologies and network capabilities have significant impacts on task migrations, especially wireless networks between computing nodes. As said in [37], network I/O performance would affect the overall application performance significantly, especially on data intensive applications which need frequent data communication. Consequently, network environment information is of particular importance to task migrations. Moreover, network condition is varied with time, so it usually needs realtime monitoring or prediction. So many researches have brought the network factor into decision making for task migrations such as in [5, 29, 46, 48].

  4. 4.

    Users’ preferences and requirements. Users should have rights to decide whether to do migrations, and migrations should satisfy users’ requirements such as the demand to finish application executions within a certain time. Users’ preferences and requirements are probably the most important factors in decision making of task migrations [46, 50].

Mobile cloud computing environment is complex, and it may be difficult to obtain all these information when making task migration decisions. As a result, getting optimal solutions may be not possible in real cloud scenarios. Instead, suboptimal solutions can be tried choosing a part of decision factors like most of proposed solutions do.

Fig. 6
figure 6

Decision factors considered for meaningful migrations

Full size image

4.5 Migration granularity

In terms of migration granularities, a finer granularity level for instance thread level or object and method level outsources computational load at a refined level, which requires intensive monitoring and synchronization mechanisms on mobile devices and cloud nodes at runtime [4, 5]. Furthermore, a finer granularity level has the issue of ensuring consistency in distributed executions of mobile applications. Higher level of granularities such as module level [52] and entire application level [28] may be more simple and have lighter overhead but may result in increased data transmission and potentially have high security threats.

  • Thread level: a thread level partition and migration[4, 5].

  • Method level: partitioning occurs at a method level and intensive methods are migrated to remote servers [27].

  • Object level: an entire object is migrated to a remote environment [44, 45].

  • Class level: objects of the same class are offloaded to the same server [28].

  • Module level: an entire module or bundle is migrated to remote processing [52].

  • Task level: a application is a collection of tasks and some tasks are migrated [4, 5].

  • Entire application: the whole application is offloaded to remote servers [28, 44].

4.6 Migration motivationa and objectives

Application or task migration mechanisms aim at augmenting processing potentials of resources constrained mobile devices in MCC and makes better executions for user tasks. For doing this, diverse objectives are considered: saving energy [27, 53] and processing time [40, 48], balancing migration incomes with overhead [46], and maximizing utilization of system resources [41, 47]. They are all for obtaining best execution performance and satisfying user requirements under various scenarios.

5 Challenges of decision making for task migration in MCC

Although there have been research efforts for task migration in MCC, it is still in early stage and there is not yet a matured approach to making this really work efficiently and effectively in a large scale deployment environment. The main challenges are:

  • Comprehensive context awareness should be achieved in order to facilitate decision making for task migrations. It is known that context awareness is particular important when it comes to mobile devices [54]. Location, weather conditions, security, and other relevant contexts should be managed besides battery, power consumption, bandwidth, application data volume, and server workload. Currently, the existing approaches on decision making for task migration consider limited contexts such as power consumption and bandwidth. However, this decision process can be complicated considering weather conditions, trustworthiness, security, and privacy of cloud nodes. This calls for a dedicated context awareness framework which supports task migrations.

  • Deep context awareness for big data computing environment is needed. Big data is arising from various applications, and it is becoming more difficult to obtain the so-called deep contexts for big data computing due to complexities of big data [56]. Deep learning has been receiving great popularity from both academics and industries due to its excellent performance in many practical problems [55], such as image and voice recognition. Deep learning may be used to help achieve deep level awareness of contexts which can be called in-depth context awareness that can help to make better decisions [56].

  • Unified standards for interacting with both mobile clients and cloud servers. Currently, every approach is developing in-house closed framework for supporting task migrations in a unique way which does not satisfy interoperability. There arises the requirement for unifying these work including framework, communication protocol, and decision-making process. Unification helps to adopt a standard way to promote MCC’s industrialization and commercialization.

  • Large-scale deployment and testing of task migrations which can be globally optimized. This is to make sure task migrations can consistently work in for example a big smart city application. The current researches are mainly tested in a small scale in laboratory. In view of this situation, we need to have a larger scale of testing for globally optimized task migrations, considering possibly conflicting optimization parameters. This also calls for more efforts for the research of light weight but powerful decision mechanisms to achieve this.

  • Adaptive and intelligent decision algorithms to make more adaptive migration decisions. Current task migration solutions in mobile cloud computing use traditional algorithms to make decisions as discussed in Sect. 3 which are limited in terms of efficiency and performance. Task migration decisions may need to combine with the latest development of artificial intelligence. Decision making can also use interdisciplinary theories and approaches such as auction or game theory in economic theories to break through the limitation of existing algorithms and obtain better migration performance.

  • Distributed decision mechanisms are needed in order to make migration more flexible and accurate. Currently, the majority of work is using a centralized way to make decisions, which may be a bottleneck when there are emergent situations or a global decision is not possible. Centralized decision making needs a powerful central decision component to consider the whole system and to communicate with all the machines and tasks. This inevitably results in heavy network cost. Distributed decision mechanisms allow tasks or machines to make migration decisions according to their own situations. Thus, it is more fine grained, flexible, and accurate. Tasks and machines can get benefit from distributed task migration decisions.

6 Conclusion

Mobile cloud computing is an important computing paradigm today. A decade of efforts by many researchers have developed core concepts, techniques, and mechanisms to provide a solid foundation for progress in this area. Task migration is the main approach in MCC which offloads tasks in small nodes to other powerful cloud nodes. Currently, there are a number of approaches to decision making for task migrations. In this paper, we present a relatively comprehensive survey on these task migration approaches and analyze their solutions to decision making. They differ from each other in application migration algorithms, granularities, decision-making factors, and decision modes (centric overall decision or distributed decision at a server level or task level). Even though we can benefit from task migrations, there are still challenges that need future efforts, which include in-deep context awareness for big data environments, distributed migration decision mechanisms and algorithms, unified stands for interactions, and other issues.