Sustainable Offloading in Mobile Cloud Computing: Algorithmic Design and Implementation Sustainable Offloading in MCC: Algorithmic Design and Implementation

2.1.1 Wireless Communication. Wireless communication technology is fundamental for enabling mobile systems and permitting SMDs to have access to the Internet; on the other hand, it is the major concern in realizing such systems due to the already-existing physical challenges for maintaining connectivity and communication quality. In mobile systems, SMDs make use of wireless communication, typically via a wireless LAN or cellular network, to attach themselves to a stationary physical network.

2.1.2 Mobility. The free, independent movement of devices consists of the fundamental nature of Mobile Computing, which refers to support network availability independent of the location, as well as the physical displacement, of devices. As a vital requirement to enable mobility, SMDs usually need to follow protocols and conduct registration procedures whenever entering new environments and areas of interest. Many protocols, such as communication handoff, have been developed as a way to minimize communication interruptions and maintain connectivity as much as possible.

2.1.3 Portability. In broad terms, portability is one of the critical factors in enabling mobility of devices. The nonattachment of devices to a stationary power supply amplifies their independent movement capabilities. Consequently, the use of batteries consists of the feature that allows mobile devices to be carried around, directly enabling, and at the same restricting, mobility. The autonomy, lifespan, or charge of batteries immediately influences mobility, as well as capabilities of SMDs. Portability unconditionally suffers from the lack of a continuous power source.

2.2.1 Smart Mobile Devices. SMDs are the original motivators for designing and implementing Mobile Cloud Computing. In the whole architecture, they are the entities with enhanced capabilities through offloading techniques, and they are responsible for the original local task execution and data storage. As described later in this study, a local execution is preferred when conditions indicate that offloading to Cloudlets or remote Clouds cannot be performed efficiently, in terms of computation gain, data storage, and energy consumption. Such conditions are dynamic and dictated by evaluations obtained through profiling toward energy and computing efficiency, which includes task patterns, computing capabilities, and the networking environment.

2.2.2 Cloudlets. Cloudlets are the unique, distinct components in the Cloudlet-based architectural model. These elements in the model perform as local resource managers, responsible for connecting local devices, executing offloading requests, and forwarding offloading tasks to remote Cloud resources in case the specific local Cloudlet capabilities are insufficient. Based on the granularity of offloading requests, Cloudlets are designed and implemented differently to cater to device-level offloading or thread-level offloading.

2.2.3 Remote Cloud. The remote Cloud primarily handles the offloading of task executions and data storage, which might be forwarded from Cloudlets or directly forwarded by mobile elements. Based on the definition presented in [98], a remote Cloud performs as a resource pool that contains clones of all mobile tasks. This resource pool can also provide sufficient “unlimited” computing capability to process different types of tasks.

2.2.4 Networking. Networking plays a crucial role in defining the feasibility of offloading in the MCC model. Mobile devices, Cloudlets, and Cloud resources present a considerable variability of capacities, features, and configurations. This diversity, when it comes to the variety of communication interfaces and means, promotes a diversity of connections among these elements.

2.3.1 Local MCC. As shown in Figure 3, in this deployment model, mobile devices in the local area initiate and maintain a wireless ad hoc network to form a Mobile Ad Hoc Cloud. SMDs can collaboratively help each other in an ad hoc manner. Some designs have envisioned offloading and task delegation into Mobile Ad Hoc Clouds [109], following the most recent trends where Mobile Edge Computing (MEC) is implemented. This approach can drastically reduce communication latency but faces more constrained scenarios due to limited SMD capabilities. Within the network, idle mobile elements can provide the computing power for others. The devices can either trigger the offloading process as resource requesters or execute offloading tasks as resource vendors.

Fig. 3. Energy-aware local offloading MCC deployment model.

2.3.2 Remote Cloud MCC. The deployment model is based on remote public Cloud resources as described in Figure 4, typically provided and maintained by public cloud service providers, such as Amazon AWS, Google Cloud Platform, and Microsoft Azure. In order to enable task offloading for mobile devices, virtual instances are established and mapped to mobile users in an always-on mode. Each virtual instance hosts a virtual clone of the specific user device, capable of receiving offloading tasks, scheduling offloading executions, and returning results via a proxy. Due to the data-center-based resource integration and energy optimization technologies, the model based on remote resources can support “unlimited” computing power and storage for the end mobile users and trivially outperform the local-based counterpart in terms of computing efficiency.

Fig. 4. Energy-aware remote offloading MCC deployment model.

2.3.3 Cloudlet MCC. This mode, as depicted in Figure 5, involves both local and remote resources. Unlike Local MCC that is established and maintained by mobile devices, the Cloudlet-based model introduces the concept of the Cloudlet as the local resource manager. The Cloudlet typically refers to a static workstation or a series of workstations that are located on the edge of local wireless networks and connected with remote Cloud resources via wired networks. By pooling the local resources and coordinating with the remote Cloud, the Cloudlet can either schedule offloading tasks locally or forward them to the remote Cloud based on the computing capability requirements and energy cost.

Fig. 5. Energy-aware cloudlet offloading MCC deployment model.

2.4.1 Task Partitioning. Offloading is performed at either the application level or device level based on the diversity of applications and the size of the system. With respect to the application requirements, tasks may involve specific-purpose equipment, which may not be available in the Cloud resources, such as GPS, camera, and other sensors. In this case, offloading cannot consider these mobile application tasks due to the lack of corresponding hardware and capabilities in the Cloudlet or remote Cloud MCC solutions. Also, for security concerns, as shown in [118] and [141], applications with sensitive personal information may not be offloaded to a third-party Cloud as a matter of preventing data loss or inter-Cloud malicious attacks.

2.4.2 Profiling. Because of the nature of mobility, the networking and computing environments may change immensely during application execution, under which the offloading efficiency may also vary largely during runtime. In order to decrease the probability that offloading overhead exceeds its benefits, during a certain interval, the execution environment information is necessarily updated as references to make the real-time offloading scheduling decision. The profiled information involves three aspects: the current task load, the computing status of mobile devices and offloading resources, and the networking between them.

2.4.3 Decision Making and Execution. This step consists of the final part in the offloading procedure, which makes the distributed task scheduling decision from a global point of view. Although each task is already partitioned, tagged, and profiled, the entire task set needs a global evaluation to optimize the executions. This optimization consists of one of the most extensively explored aspects in recent works, listed in the following sections.

• Evaluation Model

As this work will describe in the subsequent sections, which summarily enlists recent studies in Tables 2 and 3, an MCC-based offloading approach cannot suit all mobile applications; performance profoundly relies on the execution environments. However, all these works share similar offloading fundamentals: reduction of application execution time and reservation of energy can be achieved if the computational gain and the communication overhead are carefully considered and evaluated. We can then define, in general terms, energy consumption involved with local processing and data transfers and energy reservation, or gain, related to offloading.

Computational energy gain refers to the difference in computing capability between the mobile device and the offloading destination resource for the targeting task set. According to the Dynamic Voltage and Frequency Scaling (DVFS) model, as in Formula 1,

The calculation of communication overhead is even more complex due to the mix of different communication media and transmission protocols. Given a simple client/server-based offloading application, the overall communication energy overhead from the perspective of the mobile device can be defined as in Formula 3:

• Implementation Methodology

Profiled information is typically abstracted as different types of parameters in mathematical models, and the decision-making process is to find the best value in these models. Even computation-only offloading scenarios involve several parameters and factors that substantially increase their modeling complexity. In more holistic approaches, energy, computation, and mobility factors are accounted for. New designs require more sophisticated evaluation tools to implement and analyze different experiment scenarios to compare to other works. Throughout the extensive study in this survey, we could observe that the experimental-based comparison is typically oriented in two directions: simulation based and implementation based, both of which concentrate on the calculation of the task execution efficiency and device energy consumption.

In simulation-based studies, the evaluation typically deploys simulators to support the targeting experiment scenarios. The mobile simulators, such as OMNet++ [123] and ONE [75], are utilized to investigate local mobile devices and communications, while the Cloud-based simulators [25, 79] handle Cloudlets and remote Cloud properties and their respective behaviors.

On the other hand, implementation-based evaluations lack standard MCC definition, adopting different implementation strategies and system architectures. In general, enabling mobile application execution externally involves the configuration of mobile devices, offloading resources, and channeling between them. SMDs are primarily smartphones and tablets that execute Android OS [10, 34] or Windows OS [36], which can utilize JVM (Java Virtual Machine) or CLR (.Net Common Language Runtime) to solve the cross-platform issue. The lack of iOS-based Apple devices is mainly due to the insufficient support of MacOS-based virtualization on the Cloud side. The offloading resources differ based on the granularity of the corresponding topics. For instance, the Android-based image manipulation application scenario [10] is built on the Amazon public EC2 Cloud for considering and testing access to remote resources. A VM task overlay [115] is implemented on local servers in order to manipulate local resources. VirtualBox-based Android x86 image processing [81] is deployed on private and public Clouds for testing offloading approaches with hybrid resources. The communication between mobile devices and offloading resources is usually via proxies or middleware. In terms of proxy-based communication, application servers in the middle of mobile devices [81] have been used to dispatch application tasks to corresponding computing units. For the middleware-based solution, MAUI runtime is deployed inside smartphones and servers, which maintain the client proxy and server proxy internally [36]. In the context of communication, Kimberley Control Manager [115] has been adopted to provide service discovery and message transmission.

3.1.1 Availability. MCC heavily relies on the availability of Cloud services and resources due to its close dependency on mobile networks, such as WiFi and cellular communication. Besides being closely related to the existence of connectivity, availability might concern the selection of the most suitable communication channel and Cloud resources for minimizing offloading costs.

In the sense of identifying the most suitable Cloud resources, CloneCloud [34] has tackled the challenge of determining the most appropriate Cloud resource when migrating or offloading applications and tasks to the Cloud. It also considered the “distance” of the resources from the position of SMDs, defining acceptable offloading decisions based on communication latencies. Thus, CloneCloud conducted a study that emphasized the availability regarding the new MCC offloading process.

In CloneCloud, similar to MAUI [36], which utilizes the .Net framework [121], process-level virtualization is also deployed to establish an intermediate layer that can translate tasks to suit different instruction sets. Thus, mobile devices and static servers can share the same source code. Then, application tasks are suggested to be partitioned automatically based on rules, to keep application developers from partitioning details.

VD-ABC [40] deals with a similar problem as CloneCloud. It determines the optimal deployment of the cloned virtual machine in the Cloud to reduce the communication distance between SMDs and cloned VMS to minimize energy consumption. The approach employs the artificial bee colony optimization metaheuristic with Boltzmann selection polity to tackle the complexity of the problem when determining the best VM placement while acknowledging the offloading of several other SMDs. Consequently, VD-ABC must contemplate the feasibility of VM deployment in case the involved energy costs are substantially high and ultimately do not grant any benefit.

In a more restrictive perspective, EnaCloud [87] focused on the Cloud resources when performing offloading; it implemented a dynamic VM placement to enhance energy efficiency on the Cloud side. By abstracting the VM placement as a bin packing problem, EnaCloud made use of heuristics to achieve a suboptimal placement solution, where it reduced energy consumption and overhead by adjusting resource demands. Focused on the placement of Cloud VMs, EnaCloud disregarded communication costs and latencies when determining the SLA-based placement strategy.

Even though there might be suitable resources for fully implementing MCC, the high volume of offloading requests may lead to a fast lack of Cloud endpoints for receiving tasks and data. Within this context, DTMMC [60] has realized a cost analysis that concerned availability as the number of local users increases. This work first illustrates the lifespan of task offloading in the Cloud and then argues the inevitable task migration overhead brought by the unpredictable workload in such an offloading process. However, due to the lack of global-level information, hotspot servers may easily be overcommitted, which further downgrades the task offloading performance.

TIM/EDA [74] has designed an auction-based offloading within a Cloud-based MCC architecture. Resources are valued differently in this approach, where it holistically considers computation gain, communication cost, link performance, and energy consumption. The auction model considers both Cloudlets and SMDs when offloading tasks. The model aims at promoting collaboration of SMDs and Cloudlets with incentives and better fairness when hosting offloaded tasks. TIM/EDA implements a strategy for evenly distributing load and increasing the availability of resources for further offloading of other SMDs.

Finally, availability is closely related to the connectivity between devices and offloading Cloud resources, which involves coping with the instability or suitability of communication channels and interfaces. When the Cloudlet MCC model is contemplated, new communication opportunities might be available in the sense of additional communication channels, which are represented through different interfaces, and conditions, which are reflected in the proximity of resources. EaCMD [39] concerns the availability of public WiFi and proposes a collaborative offloading strategy that introduces cellular networks as offloading channels. The optimal offloading channel can be selected via channel evaluation between the WLAN and cellular network, which is performed by prefetching location-based caches. The maintenance and updates of the caches are guided by mobility prediction. Based on the comparison and prediction, the collaborative offloading strategy can achieve 80% more energy savings than the naive ones. In EeSPC [138], MCC collaboration is abstracted as a Markovian channel-based DAG, which is further handled by the LARAC protocol.

MECC [32] explores the opportunities in multichannel wireless communication. The approach considers the interference among multiple peers when attempting to access the communication medium for offloading. MECC makes use of game theory for implementing a distributed computation offloading algorithm, coping with the multichannel wireless contention. During the execution, the algorithm analyzes and identifies the benefit of offloading tasks in light of the communication conditions where transmission rates lower than a threshold do not benefit any SMDs.

EECO [136] is similar to MECC [32]; it also leads with the multichannel communication conditions and multiaccess characteristics of 5G when offloading to the remote Cloud in a Mobile Edge Computing scenario. EECO defines an optimization model to minimize the whole system energy consumption and to satisfy the latency constraints of mobile applications. Thus, the model accounts for the costs of local computation and data transmission in terms of energy consumption. The model follows a three-stage scheme for energy-efficient computation offloading where it classifies communication channels and prioritizes assignments to handle multiuser task offloading in 5G.

UMCC [96] places MCC in a much broader sense, Smart Cities, where it considers several Cloud topologies: Central Cloud, Cloudlet, and Distributed Mobile Cloud. The first objective of this approach consists of dealing with the heterogeneity of communication technologies. As a result, the model contains a utility function that considers application factors, such as latency, energy consumption, and bandwidth consumption. UMCC makes use of the Shannon Formula to solve this optimization problem where a weighted sum of offloading entities is implemented following different policies: greedy, cluster based, and biased randomization.

3.1.2 Communication Cost and Profiling. Due to the nature of the distributed MCC environment, energy-aware offloading cannot ignore communication characteristics involved with data transmissions. Availability closely relates to communication in terms of connectivity of SMDs over networks. Assuming that availability is already achieved, the cost of communication is a fundamental, delimiting factor for quantifying the actual benefit of MCC offloading.

The dynamic communication changes that might happen in mobile environments have served as a strong motivator to determine precise networking conditions as a decision-making parameter to promote MCC offloading. As a result, EACMA [3] analyzed communication overhead by evaluating the impacts of the communication distance and the number of intermediate hops. Similarly, this subject has been tackled in VMBC [115] and CIMCA [47], which suggested that WAN-level communication may compromise the usability of MCC offloading. The feasibility evaluation is also performed in terms of energy consumption and execution efficiency.

Bound to a narrower class of applications than in EACMA [3], BECMC [10] focused on determining communication overhead for dealing with data synchronization. It tested the task offloading performance of Android smartphones on the Amazon public EC2 Cloud platform [5]. In the proposed scenario, system-level information and the application code are tracked and monitored passively. Passive information is extracted through the Intents class, while Active information is driven by the instantiation of the Alarm class, which periodically triggers data collection.

Odessa [107] also analyzed the communication costs for MCC offloading, so it concerned dynamic profiling and its benefits toward adaptation during real-time offloading. This approach proposed a greedy profiler that can periodically extract piggyback environmental stage data, which is further utilized to estimate the execution bottleneck and scheduling adaptation.

Similar to BECMC [10], MAUI [36] has evaluated the energy consumption on both the data transmission and the profiling overhead. For the communication overhead, it tests a typical code uploading scenario under which a small piece of the task is sent from mobile devices to remote resources via 3G or WiFi. The results presented two positive correlations: RTT (Round-Trip Time) versus energy consumption and throughput versus energy consumption. Under a certain throughput value, 3G uploading may consume three to five times more energy than WiFi uploading. Specifically, the paper emphasizes the impacts of WiFi's Power-Save Mode (PSM). The PSM is intended to reserve energy by turning the devices into passive mode when no data is sent. However, the device state transmission cost from passive mode to active mode also introduces latency, which causes higher energy consumption for small transfers.

The restrictive communication conditions that constrained offloading in MCC models have been relaxed with Cloudlet MCC architectures. The shorter distance between SMDs and Cloudlets have allowed better offloading gains and energy reservation. CIMCA [47] has performed a transmission cost analysis study on the decision-making network, which has tested the offloading performance between mobile devices and Cloudlets in the scope of interactive offloading tasks. CIMCA achieves task offloading by one or multiple reachable Cloudlets that are peers to each other. The work also evaluates the task offloading performance based on three types of tasks with decentralized routing and centralized routing. CIMCA suggests that Cloudlet-based offloading can provide larger throughput and lower delay when compared to the remote Cloud.

As a way to counter the costly communication latencies, VMBC [115] has explored the reduction of the amount of data involved with MCC offloading. VMBC targeted the long-distance transmission overhead with the VM preparation process. The output of this work suggests that the WAN-level latency between mobile devices and Cloud resources may significantly restrict the usage of offloading applications even if sufficient bandwidth is provided. Concerning the reduction of VM state transmission overhead, the work proposed a VM overlay model to enable reduction of the size of the offloading image.

Correctly assessing the dynamic communication status of an MCC environment allows for more successful offloading decision making. Even though communication costs are not the primary objective in many offloading models, they have been recognized as critical for allowing overall execution and energy gains. EA-HRM2 [55] implemented a model that contemplates the communication costs for offloading tasks to the Cloud. In heterogeneous embedded environments, the access to the wireless communication interfaces may pose a high cost in terms of energy. Thus, the task assignment model of EA-HRM2, even though focused on computation gains, heavily relies on the energy costs involved with data transmissions.

Similar to EA-HRM2 [55], EMOP [120] developed a strategy to optimize offloading in terms of energy consumption considering the changing communication conditions within mobile environments. EMOP introduces an energy-efficient multisite offloading policy that uses Discrete Time Markov Chain (DTMC) to model fading wireless mobile channels and dynamicity of SMDs. The Markov Decision Process (MDP)-based framework formulates the multisite partitioning problem, which considers communication cost against the offloading gains concerning energy.

In the same context as both EA-HRM2 [55] and EMOP [120], eDors [65] dealt with conciliating two major dynamic offloading objectives: reducing execution time and energy consumption. The model focuses on the computation aspects of the tasks. However, to reduce energy consumption, it recognizes the communication costs for each offloaded task. Thus, eDors identifies the suboptimal tradeoff between the two objectives through Max-Min heuristics where transmission power is controlled to favor offloading depending on dynamic communication conditions.

Comparable to eDors [65], MinED [126] studied the tradeoff between energy reservation and communication delay. In this case, energy reservation is understood as primarily promoted through computation offloading. Thus, the model acknowledges that offloading promotes energy reservation by relieving local processing in SMDs. However, offloading incurs costly application delays due to inherent communication latencies. MinED then copes with this duality by formulating a joint optimization problem through 0-1 Integer Linear Programming (ILP).

Communication costs not only involve the surrounding networking environment of SMDs but also the characteristics of applications that are being offloaded. The frequency at which synchronization needs to be conducted for keeping offloaded tasks consistent dictates the communication overhead throughout the application lifespan. Consequently, SynMCC [82] introduced a model that acknowledges data synchronization for maintaining the consistency in the duality between SMDs and their respective cloned Cloud VMs. Synchronization frequency is dictated by the amount of offloaded data and changes in the data. Thus, SynMCC implements application and task profiling in SMDs for selective offloading decision making. The decision making attempts to reduce energy consumption through minimal data synchronization between SMDs and the Cloud.

Opportunism is another exploited factor in Smart City scenarios where there exist diverse network interfaces in SMDs that can extend connectivity and promote better conditions for MCC offloading. UMCC [96] considers heterogeneous urban environments for MCC. The framework focuses on the availability through the connectivity of SMDs over several Cloud topologies. However, besides maximizing connectivity, UMCC attempts to reduce the communication costs when there are many communication technologies. The framework achieves this cost minimization by using the Shannon Formula to solve the optimization problem.

Sharing the same objective of reducing the energy costs of communication as in UMCC [96], E2G [112] designed a model for efficiently transmitting data to the Cloud in an MCC system. The policy behind this approach assumes that an SMD bridging the costly communication with the remote Cloud is the most effective strategy to reduce data transmission costs of neighboring SMDs. In order to promote fairness among SMDs and avoid any selfishness, the model is based on Subgame Perfect Nash Equilibrium (SPNE). SPNE allows E2G to select an SMD responsible for bridging the communication and minimize the whole energy cost of transmissions.

The new trends in cellular networks have also enabled exploring the beneficial characteristics of 5G for enabling multichannel, multiaccess data transmissions with the purpose of boosting offloading. For instance, EECO [136] accounts for the costs of local computation and data transmission while acknowledging the multiaccess characteristics of 5G, observing Orthogonal FDMA (OFDMA). Besides concerning accessibility in 5G, the model minimizes the whole system's energy while satisfying latency constraints by considering the energy consumed in the macro and small base stations when offloading. The scheme behind the model presented a three-stage energy-efficient computation offloading through classification and priority assignment.

In the same 5G context of EECO [136], OF-MECO [135] introduced a model to maximize local communication resource allocation for Mobile Edge Computing scenarios. In a multiuser environment, the model explicitly deals with two Medium Access Control (MAC) protocols: TDMA and OFDMA. OF-MECO makes use of a threshold-based convex optimization through priorities to cope with infinite and finite Cloud capacities. On the other hand, the model formulates a mixed-integer problem to maximize channel gains against local computing energy consumption through Lagrange and Karush–Kuhn–Tucker (KKT) conditions.

Very close to EECO [136], EENA [113] implemented an energy-efficient and network-aware offloading algorithm where communication between SMDs and the Cloud is realized over 5G technology. The suggested model makes use of the 5G characteristics and contemplates the use of Cloudlets, or brokers, for connecting SMDS with the Cloud servers. EENA employs decision matrices with dynamic parameters for regulating offloading to minimize energy consumption and involved communication costs.

With technologies toward the practical implementation of Microwave Power Transfer (MPT), new possibilities can be explored to extend the battery lifespan of SMDs together with MCC strategies. MCCWET [134] focused on energy reservation through several strategies: controlling SMDs’ CPU cycles, harvesting energy through MPT, and selecting offloading. By assuming the full availability of MPT, the framework employs Convex Optimization Theory to minimize mobile energy consumption by selectively opting for offloading or controlling local power. MCCWET makes use of the threshold to interleave MPT with offloading and guarantees transmissions with optimal data allocation for dynamic channels.

CoopMCC [41] expands MCCWET [134] by collaboratively allowing SMDs to help each other in an ad hoc fashion. CoopMCC explored the cooperation among SMDs to extend the lifespan of their batteries. The approach implemented a cooperative contract where it attempts to achieve the mutual benefit of two types of mobile devices: energy-poor and data-poor. CoopMCC relies on a power-splitting (PS) policy to generate an optimization closed-form model where communication can be exchanged with energy. Energy-poor helps data-poor by relaying data, and data-poor helps and allows energy-poor to harvest its energy through power transfers. The model identifies the optimal PS factor and transmission power as a form to stimulate cooperation and exchange and achieve an optimal cooperation contract.

3.1.3 Computation Efficiency. Extending the computation capabilities of SMDs consists of a primary MCC functionality where communication costs cannot be denied due to their inherent presence in any mobile environments. With the objective of enhancing the processing of applications and preserving SMDs’ energy, offloading schemes have employed subpartitioning of tasks, scheduling based on task dependencies, and Cloud resource “limitations.”

Unlike the previous objectives that mainly concentrate on reducing overhead or coping with the availability of Cloud resources, some approaches sought to maximize the computing efficiency that MCC offloading can support. An offloading architecture has been introduced in ThinkAir [81], exploring the parallelism in the MCC-based offloading. In this proposed architecture, the application subtasks are partitioned with the ability to execute simultaneously. The offloading performance proves to be effective for computing-intensive tasks that involve little data communication during execution but large computational load.

Similar to ThinkAir [81], JACCR [8] evaluated offloading under which computation-oriented tasks are scheduled and mapped to many Cloud Computing units. This pairing enables multiple virtual instances to serve different application tasks simultaneously, increasing the level of parallelism of offloaded tasks in Cloud virtual servers. The results show that the proposed MIMO (Multi-Input Multioutput) offloading model can extensively reduce the task queueing compared to the device-only execution with the same task sets.

Following the task subpartitioning strategy and when it is possible, finer-grained tasks can lead to more flexible offloading. BCSA [93] introduced an optimal selection algorithm that makes use of adaptive task partitioning and dynamic selective offloading strategies. The primary target of BCSA consists of minimizing execution costs where local energy consumption, execution time, and execution costs are the major parameters of its search. These parameters consider the communication delay and costs, in terms of energy, involved with the offloading process.

In contrast to BCSA [93], DNN-MCC [44] explored the particular characteristics of Deep Neural Networks (DNNs) for optimizing the performance of applications when offloading them to the Cloud. Though a mobile-Cloud collaborative approach, rich DNN-based applications, such as speech recognition and intelligent personal assistant (Siri), can be deployed more effectively in SMDs. The model exploits the partitioning of heavy DNN convolutions for effectively boosting execution and reducing energy consumption, relying on data exchanges.

Due to the popularity of streaming and gaming applications in the mobile devices, some works try to augment the performance of these applications, such as MTCCO [94]. MTCCO evaluates the offloading performance in CDIA (Cloud-based Distributed Interactive Application) to traditional DIA (Distributed Interactive Application). Due to the limited CPU and GPU capabilities, the proposed scheme treats mobile devices as lightweight terminals only responsible for presenting streaming data. On the Cloud side, a mobile client is deployed to communicate with the application server on behalf of the corresponding mobile user, which forwards user commands to target proxies and routes streaming data back to the user's screen. In this case, mobile devices can leverage the gaming processing entirely to the Cloud and be in charge of only uploading instructions and showing results.

A relevant work is also shown in PEDSA [132] for data stream tasks. Concerning the maximization of the application makespan and throughput, a GA-based algorithm is proposed for the stream task partitioning, which begins with random selection and ends with maximal computing efficiency or largest throughput. The evaluation presents that using the proposed task partitioning model can always outperform the all-mobile or all-cloud naive approaches on the graph generation application. However, the computational cost of the mobile device is not fully discussed. Based on the DVFS mobile device computing model, a shorter task execution time may not necessarily lead to smaller energy consumption, due to the high computation load introduced by partitioning and channeling.

Even though task partitioning is necessary for allowing offloading to Cloud servers and achieving a certain level of parallelism of applications, scheduling decision making is fundamental for guaranteeing proper task execution to reach substantial computation gains. In that sense, CSOSM [140] introduced a client-server-based MCC system, tackling the entire task partitioning, profiling, and decision-making process. In CSOSM, the offloading scheduling process is abstracted as two aspects—offloading channel selection and resource mapping. The MCC system uses the TOPSIS offloading decision-making algorithm, which is deployed to the multicriteria wireless interface election, while the MIN-MIN heuristic is implemented on the Cloud components to process tasks. As the experiment shows, the proposed system can outperform ThinkAir [81], reducing execution time to two-thirds and preventing half of energy consumption.

In order to cope with the decision-making complexity, ACOMC [48] employed a fuzzy-logic-based code offloading engine on top of the COMET virtualization platform [90]. In ACOMC, environment profiling information is translated into crisp sets. Based on the fuzzy threshold rules, these crisp sets further generate offloading decisions. Experimental results showed offloading improvement that led to an average 10-second execution time for message delivery.

COMET [61] enabled distributed shared memory between mobile devices and the Cloud resources via Dalvik virtualization [20]. Instead of PRC, applying DSM can support flexible thread-level migration, which is presented to reach 15 times higher execution efficiency.

TaCO [64] introduced a DAG-based analysis over the offloading costs and benefits to select a task to run in either a Cloudlet or a remote Cloud server. It implements a task-centric strategy so that a finer granularity offloading model is enabled. Also, lighter application cloned VMs are used instead of the traditional SMD cloned VMs. This scheme imposes higher overhead toward low density of SMDs but compensates in saturated offloading requests.

In smart city or urban computing environments, the scale of the system adds multiuser scenarios where access introduces stress to resource sharing and application offloading. The challenge in such conditions consists of dealing with the selfishness of users and promoting fairness when allocating MCC resources for improving execution and reducing energy consumption. MCC-GNEP [26] explores “open offloading” in an unmanaged, multiuser scenario. The approach makes use of a game theory model using queuing theory to deal with the selfishness of users where the MCC system follows a Cloudlet architecture. MCC-GNEP employs a distributed algorithm for the model's architecture to achieve an equilibrium as a way to address the multiuser offloading scenario. The model assumes that applications have been already partitioned, statically or dynamically, and methodologies have determined which tasks to offload.

Similar to MCC-GNEP [26], MECC [32] considers a MEC scenario where it acknowledges multiuser computation offloading in multichannel wireless interference. Besides attempting to solve the availability of Cloud resources, the approach also uses game theory for the distributed computation offloading algorithm to cope with the selfishness of users when offloading. Also, it extends the offloading model into a multichannel wireless contention.

The distributed MCC environment inherently requires dealing with communication issues even when energy-area offloading approaches emphasize computation efficiency. Even disregarding the interdependencies of tasks, offloading depends on the networking infrastructure for sending tasks to run and retrieve their output. In this regard, EA-HCM [54] focused on task assignment into heterogeneous Cloud cores. This approach extends on a previous work, EA-HRM2 [55], where communication costs are heavily considered for the offloading decision making. EA-HCM implements a heterogeneous task assignment algorithm to reduce energy consumption involved with offloading but still considering wireless communication costs. Thus, this model solves an energy minimization problem observing heterogeneous computing and data preprocessing.

eDors [65] was listed in the previous section as an offloading strategy relying on controlling the transmission power of wireless channels to tackle high communication costs. eDors has the reduction of energy consumption and application execution time as its primary offloading objectives. Its model solves an energy-efficiency cost minimization problem in light of the task dependencies through a dynamic partition scheme over a DAG of task dependency. In the end, eDors implements a distributed algorithm based on computation offloading selection, clock frequency control, and transmission power allocation.

SynMCC [82] employs selective offloading decision making to minimize the frequency in which data is synchronized between SMDs and their respective cloned Cloud VM. Even though this approach has been listed as oriented to communication costs, its core offloading strategy relies on controlling the computation gain through different scheduling policies: periodic, programmatic, event-based, and resource-intensive update.

As described previously, for EMOP [120], although it closely considers communication costs in its model, its major focus formulates on multisite partitioning. The partitioning policy is built on a value iteration algorithm based on a state transition graph. A Markov Decision Process framework allows one to estimate mobile wireless channels for offloading tasks to multiple Cloud servers.

Most of the offloading techniques assume that applications allow partitioning in subtasks and certain level dependency on the local SMD resources. When we observe practical examples of mobile applications, we notice that they present time constraints for completing operations and returning outputs. Therefore, IDA [92] has dealt with applications’ time constraints in a multidevice scenario when offloading tasks. The model attempts to reduce the overall energy consumption of several devices in a Mobile Cloud where tasks are offloaded considering the involved cellular communication costs. The task scheduling of heterogeneous applications is mapped as a nonlinear integer programming problem where application time deadlines and transmission error rates are the major problem constraints. IDA, the interactive decoupling algorithm, is then used to combine two methods, linear relaxation and convex optimization, for reaching a scheduling solution.

In a similar context as IDA [92], EDTS [89] designed a dynamic task scheduling algorithm to minimize the energy consumption of mobile applications in light of constraints related to time and application probability. The algorithm makes use of schemes to reduce energy costs related to local processing and offloading by identifying critical paths in dependency task graphs. A suboptimal scheme solves the scheduling problem by using the Data Flow Graph Critical Path (DFGCP) and an optimal solution that addresses the Critical Path Assignment (CPA) with dynamic programming.

All described offloading techniques deal with a significant level of search complexity to achieve plausible suboptimal energy and execution gains. For the sake of simplicity, local offloading decision making is sought as an alternative to solve the energy and execution issues of SMDs in their local scope by assuming that Cloud resources are optimally available and suitable. LODCO [95] made use of a low-complexity online algorithm that implemented decision making depending only on the local mobile device's information and execution status, such as consumed CPU cycles and needed transmission power for offloading. The model minimizes the execution cost of mobile applications in the context of Mobile Edge Computing environments by using Lyapunov optimization-based dynamic computation offloading.

MAYA [83] dealt with a minimization of energy consumption and offloading costs in MCC environments. The model assumed that energy consumption was reduced through the execution of tasks in remote Cloud servers while the execution of the same offloaded tasks is posed as monetary service costs. The models had as a major parameter the classification of SMDs according to their capacity to afford homogeneous Cloud services. The energy and cost minimization is mapped as an Integer Linear Programming optimization problem to determine the best suitable task scheduling in a three-tier Cloud model.

3.2.1 Heuristics. As illustrated by the offloading approaches described in the previous sections, heuristics allows for solving scheduling and placement problems faster when compared to traditional, formal closed-form models. Usually, heuristics leads to approximate solutions that might be very close to the optimal solution of a problem, approximating the exact solution. For searching algorithms, the heuristic function allows one to rank alternatives while the search is conducted, performing as a fundamental element when defining the convergence of the search.

• Convex Programming allows for minimizing convex functions, assuming that the local minimum is the global minimum in the general case of the problem to be solved. In the context of MCC offloading, MCCWET [134] has used convex optimization to minimize the mobile energy consumption.
  

• Integer Linear Programming (ILP) “simplifies” the problem by assuming some or all variables to be integers when solving an optimization problem. Usually, the ILP objective function and constraints are linear. For solving an ILP problem, Linear Programming relaxation can be applied to relax an NP-hard problem into a polynomial solution; however, the relaxation may involve violating constraints, may not be possible, or may not be optimal. In the context of offloading, MAYA [83] uses integer linear programming to minimize energy and Cloud service cost when requesting Cloud servers to offload SMDs. MinED [126] also formulated a joint optimization problem by using 0-1 integer linear programming to minimize energy consumption and communication delay.
  

• Nonlinear Programming allows one to optimize a nonlinear, or nonconvex, problem by defining an objective function that works over a set of constraints, system equalities and inequalities, and a set of unknown real variables. The method either maximizes or minimizes the objective function, the optimized targeted objectives. IDA [92] has employed Nonlinear Programming to optimize offloading in light of the cellular communication costs.
  

• Dynamic Programming recursively breaks down a complex problem into simpler subproblems. Finding the optimal solution for the original problem consists of identifying and/or joining the optimal solution in the subproblems. EDTS [89] has used Dynamic Programming to identify the critical path in the task dependency graph and determine an optimal scheduling solution, considering time and application probability constraints.
  

3.2.2 Meta-Heuristics. Different from heuristics, which are designed to deal with the specific class of problems, meta-heuristics represent a set of “general purpose” heuristics that can be adapted, or applied, to any class of scheduling optimization problem. Evolutionary algorithms have been extensively used to search for optimization solutions. VD-ABC [40], for instance, has used the Artificial Bee Colony (ABC) algorithm with the Boltzmann selection policy to determine VM deployments as a way to minimize the communication cost with SMDs in MCC.

3.2.3 Stochastic Optimization. Stochastic optimization is broadly used to represent the random behavior of mobility and transmission uncertainty in wireless communication channels. With a random probability distribution, a stochastic model allows estimating communication conditions when they cannot be predicted accurately. EMOP [120] has employed Discrete Time Markov Chain to model wireless mobile channels.

3.2.4 Game Theory. Game Theory allows for solving conflict and cooperation problems. In the context of MCC, the cooperation of selfless users may lead to fairer split and access to Cloud resources, achieving overall higher energy preservation through offloading. MCC-GNEP [26] uses Queueing Theory to address a multiuser, unmanaged scenario and achieve an equilibrium where open offloading is implemented. MECC [32] also assumes a multiuser scenario and employs Game Theory for the distributed computation offloading algorithm that is then extended into a multichannel wireless contention environment. E2G [112] defines a model based on Subgame Perfect Nash Equilibrium (SPNE) to select an SMD to communicate with the Cloud.

3.3.1 Task Partitioning. Task partitioning needs to handle the diversity of hardware and software. The Hardware Layer issue typically refers to the CPU architecture sets; this issue requires the source code to be translated into different machine-executable bits accordingly. The Software Layer problem consists of enabling developers to migrate traditional mobile applications to MCC applications with as little effort as possible.

To tackle the usability issues, MAUI [36] seeks to minimize the effort of developers when designing offloadable applications by introducing annotations. The proposed architecture encourages developers to simply tag all classes as remote executable if they are not related to mobile-specific components and interfaces. In terms of compatibility, MAUI employs a deployment strategy to coordinate an ARM-based mobile CPU and x86-based server CPU [110]. In this case, the intermediate language can be automatically translated into different machine languages respectively, and the same source code can be executed on different platforms.

CloneCloud [34] utilizes Dalvik virtual machines [70] integrated with a Java compiler front- end to handle the compatibility. In addition, the task partitioning components can smartly tag the tasks as remote executable or local executable on the class or method level, without the developer's efforts. In this case, traditional mobile-device-based Java applications can operate seamlessly in the Mobile Cloud environment. The self-managed task partitioning limits the scope of remote tasks due to mobile-specific features and the device-VM one-on-one mapping.

Instead of implementing virtualization at the application level, ThinkAir [81] has suggested deploying an Android-based x86 virtual machine [69] directly on the Cloud. This remote deployment may introduce larger overhead while it provides higher authority that can further enable parallel task execution. Instead of automatic partitioning, ThinkAir also requires manual annotation via APIs, which may further support user-defined parallel algorithms.

In other instances, VMBC [115] also argues to utilize hardware-level virtual machines to handle the compatibility. This approach advocates that virtual instances can support a broader range of applications. Also, they are not limited by the specific languages in which applications are coded. In addition, the establishment and recycling of virtual instances are already well managed on the Cloud, which can directly support MCC applications in the same way as traditional software.

On the one hand, hardware-layer solutions outperform process-level ones in terms of the flexibility and security that virtual isolation brings. On the other hand, the virtual instances generally involve larger computing power waste. Currently, standard sample base images are still not available from public Cloud service providers. It requires smartphone designers and mobile OS developers to coordinate, define, and tailor base offloading images for different phones and tasks.

3.3.2 Profiling. The data profiling refers to computing capacity profiling and networking profiling, respectively, the results of which are the main references of the final offloading decision.

• Computing Profiling. The computing capacity profiling involves mobile CPU profiling and server computing unit profiling, both of which are performed on the DVFS energy management model. DVFS, as shown in the study described in [105], can reserve energy through delaying task execution. The model changes the voltage during runtime, and the CPU frequency can scale up and down based on the task requirements. In this case, achieving energy efficiency may not necessarily mean poor support for real-time applications. As proposed in [106], instead of calculating average execution time, the profiling model concerns the worst-case boundary and using look-ahead to reduce energy costs for embedded devices further.
  
• Networking Profiling. Unlike the computing units, the network environment, especially the wireless LAN, is not stable. Although network parameters, such as bandwidth and RTT, can be predetected before offloading, they may change drastically during execution.
  

In CloneCloud [34], the profiling data are simplified as CPU availability, screen display, and network availability. Since it assumes a one-on-one task VM mapping and ignores the energy consumption of the Cloud resources, the proposed profiling model can detect these parameters easily and minimize the overhead that the profiling process brings.

MAUI [36] classifies profiling behavior into device profiling, program profiling, and network profiling, respectively. The device profiling concentrates on the CPU utilization and predefines a linear relationship between CPU usage and energy consumption for the specific mobile device based on multiple experiments. In terms of network profiling, MAUI requires basic network parameters including RTT, bandwidth, and package loss rate. Similar to device profiling and program profiling, MAUI also deploys the test-collect method to fetch these values.

ThinkAir [81] also organizes the profiling process into three subdomains—task, hard devices, and network. The hard device profiling is straightforward, which reads the CPU usage and screen and records WiFi and 3G sensors with two states. The task profiling in ThinkAir directly utilizes standard Android APIs to document the task-related information such as execution time, required resources, uploading task size, and others, specifically tailored for Android phones and pads. For the network characteristics, it requires data regarding RTT, bandwidth, and send/receive rates that are gathered by test packets.

As can be noticed in these profiling methods, the core concern consists of the overhead that profiling itself may bring. The profiling process not only needs to occupy at least one thread to read the APIs or prestored data but also consumes energy. The solution usually comprehends to prefetch test data in advance and simplify the execution time data collection.

3.3.3 Decision Making and Execution. Based on the profiled data, the decision-making process seeks to produce a global optimization for the entire task offloading. The problem is generally mapped to the network path forwarding issue. Given a task set that involves $N$ offloadable tasks and that each task is linked to some others based on the reliant relationship, the task execution scenario can be further presented as from the first one to the $N$ task, via many possible paths. In this case, the profiled coefficient can be added into the links between tasks and the decision goal is to find the path with the lowest weighted values.

Since both MAUI [36] and CloneCloud [34] assume a centralized execution, they suggest using Shortest Path algorithms that only run in $O (Nlog (N)),$ where $N$ is the number of subtasks in a typical mobile application in which the number of edges is highly limited. Although the decision-making overhead is controlled to a certain level, it depreciates repeated scheduling and suggests to use previous results unless the profiling results changed drastically.

The issue on the decision making drastically increases in complexity when distributed offloading is enabled. In ThinkAir [81], the program profiler tags the tasks that can be executed simultaneously and calls the execution controller to arrange extra VMs to parallelize the labeled tasks. Due to the complexity of profiling and decision making, only recursive algorithms and large-scale data analysis are supported in this mode.