Importance Of Cost Benefit Analysis

Published Date: 02 Nov 2017

A classification of contemporary methods in mobile cloud computing research centered on issues related to Operational, End user , Service levels, Security & Privacy, Context awareness and Data management are presented in the report as shown in Fig. 1.[116]

Fig. 1. Classification of Methods in Mobile cloud computing.

Each issue is expanded to focus on the distinctive set of challenges faced in mobile cloud computing, and how they are addressed in existing work.

1. Operational issues

Operational issues denote the contributing technical issues such as the method of offloading, cost–benefit models that required for taking the choice of whether or not to offload, how the mobility of devices is carried out, and used connection protocols.

1.1. Method of offloading

Different research has tackled the presence of issues such as the separation of the mobile device and the cloud and the heterogeneity of the basic systems, in a variety of ways. A brief presentation of offloading methods is shown as Client–Server Communication methods, Virtualization, and Mobile agents.

Client–Server Communication.

The communication is performed between the mobile device (offloader) and surrogate device using protocols like Remote Method Invocation (RMI), Remote Procedure Calls (RPC) and Sockets. The RMI and RPC have compatible APIs and are verified by developers. The offloading by these two methods requires that the services have already been installed in the end points. It is a drawback as the ad hoc and mobile nature of a mobile cloud. This restricts the mobility of users and do not support the needed services. Spectra [27] and Chroma [28] are two systems which use pre-installed services. The functionality in remote and local Spectra servers is invoked by the applications that use RPC. Here, developers should manually partition the applications. Marinelli [14], presented ‘Hyrax’ to Android smartphone applications that are distributed in terms of computation and data established on Apache Hadoop. Hyrax examines the chance of using a cluster of mobile phones as the resource providers and shows the adaptability of such a mobile cloud.

It is indicated that the people’s location points to their activities; up on association, it leads to common activities. The Offloading manager module creates virtual machines on surrogates. On this device, the tasks are performed on a virtual machine that acts as a protected space ensuring the security. The ‘Cuckoo’ [30] discusses a system to offload mobile device applications to a cloud. It can be offloaded on any resource that has Java Virtual machine. The High Performance Programming System Ibis [31] is used for Cuckoo’s communication. A mobile version of the standard MPI over Bluetooth where mobile devices is discussed in Mobile Message Passing Interface (MMPI) framework [32] . The MMPI uses a fully inter-connected mesh structure where nodes can communicate with each other. Tasks are handled by the libraries provided in the framework. This eliminates the need for Bluetooth specific codes.

Virtual machine (VM)migration.

VMmigration means the transfer of memory image in a VM from a source server to a destination server without hindering its execution [35]. In a live migration, the memory of the VM is pre-copied without disturbing the OS; hence it looks like seamless migration. The method guarantees that no changes in code are needed while programs offload, and provide a secure execution, as the VM boundary protects the surrogate device. But, VM migration is a time-consuming and the workload can be heavy for mobile devices. Instead of connecting to a distant cloud, Satyanarayan et al. [23] suggest ‘cloudlets’. It is similar to a small data center situated at areas/places and connected to larger cloud servers through Internet.

The mobile devices are close to the resource rich cloudlet and it works as a thin client during all the resource intensive computations inside the cloudlet. The mobile device will be connected to the cloudlet using a low-latency one-hop high-bandwidth wireless connection, ensuring real time interactive response. When the user moves far from the cloudlet, the mobile device will fall back to a service mode that connects to a farer cloud serve or in worst case even work offline. Cloudlets will be decentralized, far dispersed and self-managing that requires little power, internet connectivity and control for the setup. A cloudlet only contains cached data existing elsewhere; hence loss of a cloudlet is not disastrous. MAUI [24] utilizes a mix of VM migration and code partitioning. Their goal is to conserve energy.

CloneCloud [36] also uses VM migration to offload to a server through 3G or WiFi. As they use device replicas, the mobile applications are unchanged and had no necessity of explaining methods done in MAUI [24]. They have a ‘cost model’ that studies the cost incurred in migration and execution in the cloud and relates the cost alongside a monolithic execution. MobiCloud [37] debates using cloud computing technology for MANETs (mobile ad hoc networks) in a protected way. Customary MANETs can be converted into service oriented architecture by MobiCloud. The key emphasis of MobiCloud is to deliver a security service architecture and present ‘Virtual Trusted and Provisioning Domain’ (VTaPD) is a service to manage information movements in various security areas, using programmable routing [38].

Mobile code.

The Scavenger [39] is a framework that uses cyber-foraging by WiFi for connectivity, and uses a mobile code technique to divide and allocate jobs. It also presents a scheduler for cost assessment. Its cost assessment is based on the speed of surrogate server, uses a benchmarking technique to do this. Using the framework, it is possible that a mobile device can offload to more than one surrogate and tests confirm that running the application on several surrogates in parallel is more effective in terms of performance.

Fig. 5. An overview of offloading methods used in mobile cloud computing systems.[116]

Discussion.

Apart from Hyrax [14], Virtual cloud [12] and Cuckoo [30], the recent mechanisms have made use of either VM migration or Mobile code to offload. The previously mentioned projects are built on older frameworks; Hadoop [40] and Ibis [31], that are designed for distributed and grid programming. Hence it is safe to say that the traits in this particular area bias VM migration and Mobile code over the orthodox Client–Server Communication systems. The benefits of these approaches over Client–Server Communication methods as the RPC can be specified as a reason for this. Although Client–Server Communication methods have compatible APIs and are strong, they need the applications to be pre-installed. Even the disconnected nodes are not supported in this technique. Bearing in mind the ad-hoc nature of mobile systems, this is a drawback. Furthermore, the uninterrupted on-going communication between the client and server will lead to network congestion.

VM migration is used by a many of the frameworks, counting MAUI [24], Cloudlets [23], CloneCloud [36], and MobiCloud [37]. Virtualization prominently reduces the load on the programmer, since a very little or no rewriting of applications is necessary. Nevertheless, full virtualization with automatic partitioning is improbable to produce the similar fine grained optimizations like hand coded applications, while rewriting each and every application to code offload is not practical. MAUI [24] does not depend up on pure VM migration like CloneCloud [36] and Cloudlets [23], but utilizes the mix of VM migration and programmatic partitioning. Nevertheless, in situations where the mobile device is in the range of a surrogate device for a brief period, Utilizing VMmigration can be very heavy, as is shown out in [39] that use mobile agents for its appropriateness in a dynamic mobile environment.

Comparing the evaluation results from some projects against each other is difficult as the performance and energy savings are interlinked to the application. It is to be noted that even while using the same framework, performance changes for different applications. The size of the input and connection protocol type (whether 3G or WiFi) also plays a major role.

1.2. Cost–benefit analysis

It is significant to analyze the costs of offloading on to the cloud such as time, energy and monetary, versus gigantic execution/storage in advance. Walker et al. [41] discusses a model for evaluating the benefits of hiring storage from a cloud service going against buying hard drives, considers factors such as ,cost of hard disks, cost of electricity, cloud storage price per GB, disk power consumption, expected storage requirement, and human operator salary.

Li et al. in [42] suggest a model with a collection of metrics to compute the cloud cost. They consider two important costs of cloud computing: Total Cost of Ownership (TCO) and Utilization Cost. The Total Cost of Ownership (TCO) is used as the financial estimate to estimate the costs towards owning and managing an IT infrastructure. In cloud computing, TCO is believed suitable to function for providing an estimate to the commercial value of cloud venture, and considers the costs of server, software, network, support and maintenance, power, cooling, facilities, and real-estate. Utilization Cost denotes to the actual resources being used by a particular end point or application as per the dynamic demand. Instead of statically accounting the cash outlays, cloud cost analysis must include the impact of elastic utilization.

Even though some of the issues like power consumption and cloud pricing are related to mobile cloud computing, additional issues and perspectives should also to be taken in to account.

Importance of cost–benefit analysis.

In a mobile cloud, due to the essential mobile accounted dynamic nature, the resources will mostly change in any given moment. Hence, a cost–benefit analysis is necessary to weigh the profits of offloading against the probable gain by assessing the predicted cost of implementation with user specific necessities.

Cost models using resource monitoring and profiling.

Spectra [27] and Chroma [28] are two cyber-foraging systems to mobile devices that use methods to evaluate the cost vs. the benefit of offloading with surrogate servers. To evaluate the best trade-off, the gain also needs to be foreseen. So, Spectra makes use of previous work done by Narayanan et al. [43] that use a prediction model built on the notion that a resource consumption of an process is comparable to recent executions of related operations. Chroma uses a rather related method called ‘tactics’ indicated in a declarative semantic, while constructing on ideas from the earlier works of Odyssey [44].

The Scavenger [39] framework employs a cost valuation method to decide if offloading should be done or not. This is carried out by the ‘scheduler’ component that considers the factors: Relative speed and current utilization of the surrogates, Network bandwidth and latency to the surrogates, Task complexity and Input and output size. MAUI [24] uses a cost–benefit analysis by describing each technique in an application in an order. Recordings of network bandwidth and latency are also noted to include into the cost. Explicitly, MAUI’s describer takes three issues into account the devices’ energy use, Application characteristics and Network characteristics. The data from the describer as mentioned above is then given to the MAUI ‘Solver’ to select if a method should be performed locally or remotely. The Solver attempts to give the best possible dividing strategy that will give the minimum quantity of phone battery consumption.

Clonecloud [36] uses a ‘Dynamic Profiler’ to gather data used in the cost–benefit analysis, that is then fed in to the ‘Optimization Solver’ to choose which approaches needs to be migrated, so that the cost of migration and execution will be reduced. Here, the cost could mean the execution time, energy consumption or resource footprint. In [46] Zhang et al. take four attributes into account when calculating the cost of migrating mobile apps in to the cloud: monetary cost, performance attributes, power consumption, and security and privacy. These are concluded from various sensing modules in the mobile device and cloud that observe data such as network, device loads, battery, cloud loads and latency. After treating these inputs, the cost model chooses on a suitable course of action such as, migrating apps to the cloud/mobile device, choosing between different networks and assigning cloud resources.

Cost models using parametric analysis.

In [47], Kumar and Lu deliver an analytical model for matching energy norm in the cloud and the mobile device. The model takes the following factors into consideration; the speeds of the mobile device (M) and the remote cloud (S), the number of instructions of the computation (C), the number of bytes to be transferred (D), network bandwidth (B), the energy used by the mobile device in idle (Pi), computing (Pc) and communicating (Ptr) states. Supposing the cloud is F times faster than the mobile device; they deduce the amount of energy saving to be given by the Formula.

(1) If the formula gives a value more than zero, an energy conserving is possible, i.e. DB should be less when compared to CM and F should be large. So, using this model, offloading is advantageous in cases where substantial computation is needed with somewhat low amounts of communication. Wang and Li in [48] identify four kinds of cost factors; Data communication cost, Computation cost, Task scheduling cost and Data registration cost. These costs are stated as functions of runtime factors such as buffer size, input size and command line options. These are then used in their partitioning algorithm to decide an efficient partitioning subjected to the parameters.

Cost models using stochastic methods.

The mobile cloud service in the model given in MobiCloud [37] Liang et al. [49] recommends an economic mobile cloud computing model based on Semi-Markov Decision Process (SMDP) [50] for resource distribution. MobiCloud describes a system in which mobile devices use application components named ‘weblets’ that can be either transferred to the cloud, or run on the mobile device. The SMDP odel is depended on three states in the mobile cloud; a new weblet request, intra-domain transfer weblet request, and when the weblet leaves domain. As the cloud receives a request for transfer from a mobile device, it only accepts it if there can be an overall system gain. The system gain is based on maximizing of the cloud profit and reducing the expenses of the mobile user. The expenses of the mobile users depends mainly on the trade-offs of energy consumptions on a mobile device vs. the monetary costs of offloading in to the cloud. Besides the small gain, the overall system gain also considers the CPU cost in the cloud server because of the virtual image occupation. A ‘reward model’ is used to compute the costs based on the systems’ state and its action.

Discussion.

The cost models present in current mobile cloud computing systems, mainly belong to three categories: history based profiling, parametric, and stochastic. An overview of these are given in Table 1.[116]

Spectra and Chroma are quite similar, with Chroma being developed from lessons learned from Spectra. Several later works created on concepts from these two, while adding new techniques to address their drawbacks. For instance, the cost models of MAUI are based on the consideration that past invocations of the same, or a comparable operation is a good sign of its current resource usage. Nevertheless, the energy consumption of an operation in MAUI is stated as a function of the number of CPU cycles it needs, while in Spectra, energy measurements are directly used from the mobile device’s battery. Scavenger also registers history data and retains profiles similar to Spectra, but improves the concept by recording dual profiles per task. While a detailed cost–benefit analysis is essential for optimal performance, the cost of cost analysis itself should not be disregarded. This issue is deliberated in MAUI, where the overhead of execution of frequent profiling and precise estimations based on latest data, are well-adjusted to give a positive outcome.

1.3. Mobility management

One of the key issues faced in a mobile cloud is the design of intelligent mobility management techniques that provide user mobility while providing a seamless service. Defining a device’s current location is helpful to measure its potential to move away from or to the active mobile cloud. Work on localization predominantly falls into either infrastructure based methods, or peer-based methods. Infrastructure based methods use technologies such as GSM, WIFi, ultra-sound with RF, GPS, RFID, and IR. Additionally, these methods are energy intense and hardly suit the traditional needs of mobile cloud devices. In distinction peer-based techniques are better fit to manage the mobility of contributing devices, bearing in mind that relative location information is sufficient, and most can be employed with short range low power protocols like Bluetooth.

Peer based techniques for determining the position of a mobile device.

In ‘Escort’ [52], a human localization system implementing Mobile Phones is offered. Without using GPS or WiFi, the location of a person is concluded using social encounters amongst users via audio signaling, and monitoring the walking personas of dissimilar individuals via phone compasses and accelerometers. Not using GPS or WiFi that are both battery draining approaches, is particularly useful as well. In the case of ‘Escort’, fixed beacon transmitters located at random locations are used to correct mistakes in routing, since they deal with equally long distance paths. In ‘Virtual Compass’ [53], short range protocols like Bluetooth and WiFi are used to create a two dimensional demonstration of nearby devices. Peer-to-peer messaging is used to approximation the distance via signal strength, and to pass details about each device’s neighbors and the distances. A clarification exists in the method followed in ‘Virtual Compass’, there, it employs a self-adaptive scanning method that controls its scanning intervals by keeping track of changes in its neighbor graph. This regulation makes use of a central server however, where preferably a decentralized solution is desirable in the framework of mobile cloud. One such decentralized operation is ‘Friends Radar’ [54], which uses location updates in peer-to-peer fashion using XMPP. Friends Radar varies from the previously introduced systems in that only ‘known’ contacts, or friends’ locations are evident, and that it uses GPS. This method fails indoor. Similar peer-based localization techniques such as NearMe [55], and Beep Beep [56] also exist, but these do not perform at all for more than two nodes. DOLPHIN [57] require to have singular ultrasound hardware that cannot be expected to be on a normal phone.

Managing mobility via fault tolerance methods.

In [12], the location and the quantity of surrogate devices are significant since the cloud’s task is to provide maintenance for users with similar goals. Also, the number of devices in the area is needed in the measuring of applications. Using an ad hoc discovery technique, a p2p component measures the resource pool, and notifies the context manage if there is any change. Furthermore, mobility tracking is a significant part of fault tolerance. If an unstable node could be recognized beforehand, the system could take safety measures by stimulating task redundancy. Mobility is a major reason in mobile clouds for disconnection, similar to hardware malfunction in a distributed system. In Hyrax [14] disconnection starting from the mobile nature of devices is controlled through the fault tolerance mechanisms of Hadoop. As given in [40], one of the main suppositions of Hadoop and the HDFS architecture is that ‘hardware failure is common’. In [58], the accessibility of a mobile device as a resource provider is measured by its mobility. Hence, devices/users that are highly mobile are termed as less reliable as they are prone to disconnection.

Supporting mobility via component and proxy migration.

MoCA (Mobile Collaboration Architecture) [59], is a middleware for association on context-aware mobile devices. MoCA currently works with 802.11 and follows the client–server model with a framework for applying application proxies and basic services for cooperative applications. Here, user mobility is supported by observing the locations of the users and changing to an application proxy more suitable to the new location. The ‘closest proxy’ is estimated by the location of the mobile device which is inferred employing radio frequency signal pattern as implemented in project RADAR [60]. Hydra [61] facilitates emerging distributed mobile applications in an inescapable environment by the construction of a virtual computer through the involvement of a networked set of inescapable computers so that the application fulfills a mobile user’s prerequisite changes based on location, current tasks and number of people. Using RFID, spatial regions like parts of a room or a building can be recognized within a meter. The locations of objects (mobile devices) are recognized by identifying these spatial regions that contain them. In [62], a mobile service cloud based on an overlay-based scattered infrastructure is presented. This is an addition of previous work in Service Clouds [63] where an overlay-based network necessities dynamic and on demand prototyping and distribution of services. Here, mobile devices connect to overlay hosts who perform services on the wireless edge. User mobility is handled by moving the proxy service to different locations following the user.

Discussion.

A majority of mobile clouds provide mobility through component and proxy migration. Although this can be used for mobile devices linking to remote servers, it is not compatible mechanism in these cases; The mobile devices are resource providers and are commuting in ad-hoc manner, and in the mobile device offloads works to a local resource provider like a cloudlet. A probable answer for the cloudlet model is to use the same method as in ‘Follow Me’ [64], a localized and distributed location sharing system using Plug Computers as Bluetooth scanners. As stated in [12] keeping track of other mobile resources moving along with the client to form ‘communities’ is the only answer suggested so far in a mobile cloud of this type, and even is not yet been fully implemented.

1.4. Connection protocols

The current mobile cloud computing research uses a variety of connection protocols for communication including WiFi, Bluetooth, and 3G, though the majority has employed WiFi for many reasons.

WiFi (wireless Ethernet 802.11b) and Bluetooth both work in the unlicensed 2.4 GHz ISM band. WiFi was initially intended as replacement cabling for resource and peripheral sharing (such as printers, shared storage devices) among PCs, terminals etc. for wireless local area networks (WLANs). WiFi has a longer range, with radius 100m and supporting 11 Mbps data rates.

Bluetooth on the other hand, was projected to nonresident equipment and devices as wireless headsets etc., wireless personal area network (WPAN), and characterized by low power requirements and low-cost transceiver chips [65]. The range of Bluetooth typically is 10 m, depending on class, power, and physical hurdles in environment. Nevertheless, according to Bluetooth specifications, upcoming versions will be quicker up to 24 Mbps and consume less energy.

3G (third generation mobile telecommunications) is for mobile service and it shares the basic business model with telecommunications services model. The infrastructure is owned and managed by service provider and sold to customers typically on a monthly usage basis. Although the focus of cellular technology has been voice telephony, data services has also started to attract attention. Mobile broadband access of several Mbps is available via recent 3G releases such as 3.5G and 3.75G [66], although this is substantially lower than the data rate of WiFi.

Experimental results. Based on the experimental results presented in related research on mobile clouds, the energy consumption of 3G is shown to be higher than WiFi [24], though data for similar statistics for Bluetooth exists. In MAUI [24], the mobile phone using 3G to offload work to a remote server consumed three times as much energy as WiFi with a 50 ms RTT, and five times the energy of WiFi with a 25 ms RTT, meaning that downloading a 100 kB file repeatedly over 3G will deplete the battery in less than two hours. In CloneCloud [36], experiments conducted on three applications with WiFi displayed a latency of 69 ms and bandwidth of 6.6 Mbps, while offloading with 3G resulted in a latency of 680 ms, and bandwidth of 0.4 Mbps. Concerning speedups, WiFi gave speedups of 12×, 20×, and 10× while the energy consumption was 12×, 20×, and 9× less energy than the monolithic application. However, test results of 3G offloading gave lower gains; 7×, 16×, and 5× speed-up, and 6×, 14×, and 4× less energy for the same applications respectively. Greater latency and lower bandwidth of 3G is given as the cause of this.

2. End user issues

End users issues relate to issues that directly involve users such as incentives for participating, interoperability and cost. In particular, we hope to answer the following questions in this section; in what ways would users of mobile cloud services be billed? In cases of collaborative mobile clouds, how is credit represented? And how can users’ are persuaded to contribute to the resource cloud? What are the presentation and usability issues that need to be addressed for mobile cloud services?

2.1. Incentives to collaborate

In cases of mobile devices’ themselves acting as resource providers as discussed in Section 3.2 and in works such as Hyrax [14], the participating devices need to have incentives as to ‘loaning’ their resources. In [12], users are enticed to participate in sharing their mobile resources by ‘common goals’. If many users need to execute the same task, it can be partitioned so that each user only has to do a small part. The result of the task will be shared among all the participants. For the same kind of collaborative cloud computing, monetary incentives can also be considered in the means of micropayment schemes as discussed in [69], [70] and [71]. Their focus is on message transmissions in delay tolerant networks (DTN) formed by typical mobile devices. Thus, a selfish mobile host will only relay a message with a certain priority or higher since it implies the sender is willing to pay a price for successful delivery. The social incentives are based on the premise that even a selfish host will have a set of social relationships, and hence, will not display the same behavior towards all the other hosts. Other methods include enforcement schemes employed in peer-to-peer file sharing systems to control free riding [72].

2.2. Presentation and usability issues

Although there is lack of focus in this issue in mobile cloud computing research, presentation in the user interface does pose a valid challenge. This has been frequently discussed in mobile computing research [73–76], to design and develop separate user interfaces (UIs) for each and every type of device would be highly inconvenient and unrealistic for the UI developers [74]. Furthermore, user interaction methods with the application may depend on the user’s context such as location, bandwidth and remaining energy [76]. In addition, users of mobile cloud computing frameworks would also need some user level controls in the interface specific mobile clouds

3. Service and application level issues

Service and application level issues relate to the factors concerned with performance measurements of the system, and the QoS of the system. For example, in what ways do mobile cloud computing systems ensure availability? What are the faulttolerance (FT) mechanisms employed to ensure smooth execution and uninterrupted service? Cloud APIs providing libraries to support cloud application development for mobiles are also discussed.

3.1. Fault-tolerance for meeting availability requirements

Fault-tolerance is a highly important aspect in a mobile cloud, even more so than a conventional cloud because of the mobile nature of the devices, i.e. ‘‘mobility is inherently hazardous’’ [2]. Disconnection can happen due to user mobility as devices enter and leave a network. Running out of battery power, network signal loss, or hardware failures are other common factors such as Redundancy, Proxy migration and Resource tracking.

3.2. Supporting performance at service level

A majority of Application Programming Interfaces used to build mobile applications targeted for cloud computing are based on service oriented architecture such as REST and/or SOAP. Mobile applications are able to connect to and request services hosted on a remote cloud through interfaces. However, mobile Web services need to consider additional constraints other than standard Web services: frequent loss of connectivity, low computational resources, and low bandwidth.

3.3. Cloud APIs

Mobile clouds have been implemented using APIs provided by distributed computing frameworks such as Hadoop [14] and Ibis [77]. Futhermore, there are cloud APIs catering to mobile devices as well. For example, the Funambol Cloud API13 provides server and client side SDKs to develop mobile cloud applications and services that make use of images, calendar, contacts etc. stored in a Funambol server. Other open source APIs include Eucalyptus,14 Nimbus,15 and OpenNebula.16 Commercial cloud APIs include frameworks such as Dropbox,17 Azure,18 Amazon and Google Apps.

4. Privacy, security and trust

Whether offloading intensive computations, or data storage, using the cloud for mobile devices’ does pose questions of security and trust issues. In her article in [80], Kharif outlines the potential pitfalls in using cloud services for mobile devices. Because of the low capacity of mobile device storage, many users are starting to store data such as contacts, calendars and SMS on clouds. Is cloud security applicable to mobile clouds as well? Mobile cloud computing inherits the security threats of conventional cloud computing in cases when the definition of mobile cloud means to connect mobile devices to a remote cloud. In this case, the remote cloud server would be the same as a conventional cloud computing provider, making the general cloud security threats valid. At the same time, mobile clouds present a group of issues that are particular to mobile devices offloading jobs through wireless communication channels. Furthermore, security concerns that are specific to mobile devices such as battery exhaustion attacks [82], mobile botnets and targeted attacks [83] should also be considered.

4.1. General cloud security

In [84] Brodkin outlines seven security risks users need to consider in Cloud computing;

1. Privileged user access: offloading sensitive data to the cloud would mean the loss of direct physical, logical and personnel control over the data.

2. Regulatory compliance: the cloud service providers should be willing to undergo external audits and security certifications.

3. Data location: the exact physical location of user’s data is not transparent, which may lead to confusion on specific jurisdictions and commitments on local privacy requirements.

4. Data segregation: since cloud data is usually stored in a shared space, it is important each user’s data is separated from others with efficient encryption schemes.

5. Recovery: it is imperative that cloud providers provide proper recovery mechanisms for data and services in case of technological failure or other disaster.

6. Investigative support: since logging and data for multiple customers may be co-located, inappropriate or illegal activity should they occur may be very hard to investigate.

7. Long-term viability: assurance that users’ data would be safe and accessible even if the cloud company itself goes out of business.

4.2. Mobile cloud security

In addition to the aforementioned concerns, securing a mobile cloud introduces the following challenges as discussed in [85] where the authors propose a security model for elastic applications made up of ‘weblets’ that can be migrated to and from a cloud to a mobile device:

1. authentication between the weblets that would be distributed between the cloud and the device,

2. authorization for weblets that could be executing on relatively untrusted cloud environments to access sensitive user data, and

3. establishment and verification of trusted weblet execution cloud nodes.

Their security framework is based on the assumption that the cloud elasticity service (CES), including the cloud manager, application manager, cloud node manager, and cloud fabric interface (CFI), is trustworthy. The security threats are categorized as threats to mobile devices, threats to cloud platform and application container, and threats to communication channels. The authors propose a framework with the following security objectives: Trustworthy weblet containers (VMs) on both device and cloud, authentication and secure session management needed for secure communication between weblets and multiple instantiation concurrently, authorization and access control enforcing weblets on the cloud to have the lowest privileges, and logging and auditing of weblets. MobiCloud [37] aims to provide a security services architecture for MANET clouds in three ways:

1. Acting as an intermediary for identity, key, and secure data access policy management.

2. Protect information belonging to mobile users by means of security isolations.

3. Assess risks by monitoring MANET status.

4.3. Privacy

As the recent incident regarding CarrierIQ being installed and collecting information from mobile phones [86] shows, it is important for mobile phone users to have transparency and choice. It is vital that any personal data that is shared is done so with users consent and that they can choose to opt out of any data collecting program at any time. In [87], Fahramair et al. present the following requirements of a mobile and ubiquitous system that satisfies user privacy: protection against misuse, identification of pirated datasets, adjustment of laws (to provide additional security under certain circumstances), and ease of use. These are valid requirements for a mobile cloud as well. Techniques of anonymous routing such as onion routing can also be used to provide privacy for mobile nodes in a decentralized mobile cloud. Examples exist in the p2p domain, such as [89–91]. However, there are certain overheads and a risk of unreliable delivery associated with most anonymous p2p routing protocols [92]. As a solution, the degree of anonymity should be flexible and depend on the context. In addition to an authorization scheme, users of the mobile cloud should also have the ability to change their privacy settings and dictate what information can be seen.

5. Context-awareness

Schilit et al. [94] describes the three important aspects of context as: the user’s location, other users in the vicinity, and the resources in the user’s environment. For example, in a mobile user’s perspective, ‘context’ means things such as lighting, noise level, network connectivity, communication costs, communication bandwidth, and even the social situation.

Significance of context-awareness for mobile clouds.

Systems with context-awareness are able to use contextual information to change and automatically reconfigure their configurations to adapt to the context [44]. This behavior is very useful in the case of mobile systems since these deals with an execution environment that is subject to constant change. In the case of mobile cloud computing, context awareness can be used in forming resource clouds as well as processing information. For example, a device can infer its location through GPS, Bluetooth, or some other forms of positioning and use that information to prepare itself for upcoming processing.

In the rest of this section, we review the use of context awareness for mobile cloud computing systems, and also discuss a key concern of mobile device, energy awareness.

5.1. Context-aware service provisioning

It has been suggested that, mobile clouds can utilize the sensing abilities of their mobile devices such as location, acceleration, etc. and act as providers of context awareness/information. In [95] the authors suggest utilizing the sensing capabilities of mobile Internet devices to provide such context-aware service provisioning. Consider a mobile device connected to a remote cloud service through the Internet. As the context of the user changes, this prompts invocation of different cloud services based on the current context. With this kind of context-awareness, a service would not be bound to a user. Instead, when a mobile user invokes a cloud service, the request is accompanied by his/her context information, and the most suitable service is selected based on that information. Therefore, context is used to provide personalized services, and also as fault tolerant mechanisms such as rectifying low quality of service (QoS) problems. Here, the authors identify a model consisting of four layers of context elements:

1. Monitored context: monitoring device context which includes the environmental and device settings.

2. Types of gaps: refers to gaps that happen as a result of content changes.

3. Types of causes: refers to the factors that can cause the gaps.

4. Adapters: refers to the remedial actions that should be taken to remove the causes.

Volare [96] introduces a middleware for monitoring the context of a mobile device that is connected to a cloud service, and dynamically adapts the services so as to make them more resource efficient, reliable and cost efficient. Depending on this contextual information, VOLARE tries to adapt each service request by comparing the current QoS level with predefined thresholds. If at any time, the QoS level and the cost changes beyond the predefined values, VOLARE will automatically rebind to another service that can satisfy the requirements. In MoCA [59], a proxy can register an ‘interest expression’ on a mobile. These context specific ‘‘interests’’ depend on those specific applications requirements. Using ‘Intelligent access’ for Mobile Cloud Computing is discussed in [97], where use of context information provided by terminals, network nodes, or sensors positioned in the users environment enables efficient network access management across different Radio Access Technologies (RATs). Mobile cloud computing requires a wireless connection with a set of different necessities than classical heterogeneous access scenarios: connectivity for long periods, scalable bandwidth, network selection and usage based on energy costs. To satisfy these requirements, the paper proposes Intelligent Radio Network Access (IRNA). Their proposed context management architecture (CMA), based on the producer–consumer role model such as given in [98], is responsible for acquiring, processing, managing, and delivering context information. To control the supply of context information according to the mobile cloud’s requirements, the framework has a Context Quality Enabler (CQE). The CMA is made up of three components Context Provider (CP), Context Broker (CB) and Context Consumer.

5.2. Risk assessment using context-awareness

In MobiCloud [37], context information is used to facilitate risk assessment and routing decisions. MobiCloud introduces Virtual Trusted and Provisioning Domains (VTaPDs), VTaPDs identify these separate flows and create virtual domains. By doing this, a user is able to securely run multiple applications on different security domains, and to separate services for different settings based on context. Risk management is aided through context information because the status of the entire system is available. From this centralized data collection and processing, knowledge of the full MANET system is gained, and MobiCloud can easily identify malicious nodes.

5.3. Identifying potential resources and common activities using context-awareness

In [12], a Context Manager component is positioned to sense context information and store it to be used for other components. Location and number of nearby devices in the vicinity are the key contexts, with location information used for mobility traces, and number of devices used to aid the forming of the mobile cloud. Thus, the system is made aware if a new device enters the resource pool, or leaves it, thereby leading to better scalability and content distribution.

5.4. Energy awareness

Because a mobile device operates on a finite supply of energy contained in its battery, energy is one of the key resources that has to be used carefully [99]. In the context of mobile clouds, the cost of participation (such as power consumption) should be less than the benefit gained [100]. For this reason, being aware of a device’s energy usage is vital. In the following we discuss the research on energy consumption, mainly focusing on work on energy profiling and energy usage estimation.

Energy profiling.

In PowerScope [102], the authors present a profiling tool for mobile applications. The tool maps power usage to specific code components in applications and the operating system, allowing an analysis of power draining procedures. Profiling is done offline after collecting data, to ensure no overheads are added to the analysis. A digital multimeter measures the electric current used by the profiling computer and the energy profile is generated using these correlated current measurements. In [103], Rice and Hay present a power consumption measurement framework, specifically for mobile phones. Power consumption is measured by sampling the voltage drop across the phone battery and a high precision resistor. The work discussed previously measures the energy via hardware. A different approach is to take the measurements via software to query battery levels as done in PowerSpy [104], In the event tracking stage, the application is run and tracked for CPU time, I/O activity and energy consumption. In the analysis stage, the data acquired in the previous stage is processed. Work done by Cano et al. [105,106] provides insight into the energy consumption of the Bluetooth protocol. The focus of their work is on the different states of Bluetooth.

Energy usage estimation.

The exact formulas for calculating the estimation are given in [107]. Their methodology gives an Energy Cost Framework as follows:

• Overall Energy cost = Computational energy cost + Communicational energy.

• System Energy cost = Overall Energy cost + Overall infrastructure energy cost.

5.5. Discussion

Context-awareness has been utilized in mobile cloud frameworks for several tasks, including service provisioning with a better QoS, risk assessment, identifying resources and common activities. A majority of the works discussed use context information to provide a better service by personalizing the services, and providing fault tolerance through context such as user preferences, location, current QoS, network bandwidth, battery consumption, and CPU usage.

Nevertheless, the performance and energy costs of employing the sensing capabilities and the trade-offs with benefits gained have not been discussed. Although certain contextual information such as monitoring battery consumption will not add an extra toll, other information such as location monitoring could prove to be too expensive.