Towards a collaborative context-aware offloading scheme in mobile cloud computing

The use of modern mobile devices over traditional desktop and tab devices is dramatically increased. Mobile devices have limited resources regarding computation power, storage and battery. Moreover, mobile apps market is a huge market with billions of applications. Some of those applications like gaming and pattern recognition are attracted users to be installed. This kind of apps are known as resource-hungry apps due to the need for a powerful device to normally run such apps. From here, researchers work to present efficient solutions to augment the resource-poor mobile devices to be more powerful. One of those solutions is to collaborate among neighboring devices to offload some part(s) of an application. Mobile device cloud (MDC) comes with this idea in order to save time and energy. Moreover, new studies aim to investigate for social factors among mobile devices. Devices sharing some kind of friendship or some common interests are more likely to meet and exchange information. We study the social factors in a comprehensive way in order to see which factor or combination of factors are the best in conserving execution time and energy. In our investigation study we exploit for some of well-known tracefiles (like Sigcomm09 and Unical14). For those tracefiles we do sanity check and clean the datasets form noise. Then, we define a set of connectivity metrics for a pair of nodes, including: number of contacts, duration of a contact and intercontact time. After that, we quantify for the datasets observations to know which strategy is the best in terms of the number of contacts and total gain acquired by applying the connectivity metrics. The output of this study indicates that a pair of nodes with at least 4 common interests have the highest number of contacts with about 63% among all contacts. Moreover, adding direct friendship to the pair with at least 4 common interest gives the highest gain regarding the connectivity metrics. with about 64%. Thus, we think that exploiting for the two outputs from this study will give best results in the case of task offloading in a way to save time and energy. Furthermore, we test for our investigation study numerically against some of proposed offloading algorithm in order to get better understanding of the effects of social factors. Here we use some profiles based on real devices and consider some task capabilities based on real test-beds. Moreover, we design a set of offloading algorithms one of them is based on our investigation study (S-based) and the other algorithms exploit for one social factors. Indeed, we test for the replication factor on the offloading performance. Furthermore, we test for medium and high computation tasks against local execution and against our proposed algorithms to see which strategy is the best in terms of number of contact and the gain regarding the connectivity metrics. This help us in answering 2 questions: (i) when to offload a task?, and (ii) what kind of tasks is good to be offloaded?. Numerically, the algorithm based on our investigation study works fine and gives good results regarding time and success rate. The same algorithm saves more than 65% of time regarding local execution and more than 40% regarding Random offloading. Moreover, S-based gives results close to the lower-bound of the offloading (i.e. Flooding). Respect to replication factor, only 2 replicas are enough to reach the maximum performance and to overcome for tasks loss or failure problems. Finally, we run the whole set of algorithms in real simulation environment using the ONE simulator. Here, we test for number of hosts equivalent to Sigcomm09 (i.e. 76 hosts). In the ONE simulator, we propose that around one-third of the nodes are offloaders each with u tasks and the other hosts are offloadees. Moreover, we propose two cases: the first one considering all devices of the same profile (homogeneous case) while the other consider the offloadee profiles are higher (heterogeneous case). Simulation results under the ONE simulator are very close to the results obtained by numerical simulation.