An adaptive offloading decision scheme in two-class mobile edge computing systems

With the huge growth of data exchange and the increasing number of connected devices to the Internet, Mobile Cloud Computing (MCC) paradigm with its centralized approach will face tremendous loads. Therefore; the European Telecommunication Standard Institute (ETSI) came up with a new idea to overcome the problem of latency mainly and decrease the consumed energy of transmission. This new approach, Mobile Edge Computing (MEC), is proposed to use the capacity at the edge of the network such as Base Station (BS) or Access Point (AP). In this way, the amount of sent data to the cloud will be significantly reduced by having the edge server executing all the offloaded tasks on behalf of mobile users. In mobile edge computing, the system can be modelled as a single-class MEC system which considers all the tasks as the same (no differentiation in terms of priorities), or a multi-class system. In a multi-class system, the tasks originating from different mobiles may have different priorities. In case of a two-class system, there will be high and low priorities. In this study, we investigate the offloading problem in the presence of two user classes: one is high priority and the other is low priority. The tasks are generated by mobile users and are offloaded to the edge server to be executed. We assume that the server is in charge of making offloading decisions. A task is decided to be offloaded or not based on the comparison between sojourn time in case of local execution (at the mobile device) and sojourn time in case of remote execution (at the edge server). The task is offloaded if the sojourn time in case of remote execution is smaller. We assume that the edge server employs weighted Round-Robin (WRR) processor scheduling, which can be modeled as a priority Proessor Sharing (PS) queue if the time slots are considered to be small. (WRR) uses the same principle of Round Robin, which is basically sharing the CPU service among all the packets for a specific number of time slots (Machine Cycles (MC) in case of CPU scheduling) without considering the priorities of arrived tasks. On the other hand, WRR takes into account the priorities of the tasks. In WRR, a high priority task is served more frequently than others. To put it another way, a high priority task receives more MCs than lower priority tasks. It is known that the expected sojourn time in a PS queue is linearly proportional to the task size. Therefore; we use a factor (queueing delay multiplier) that accurately captures the sojourn time of an offloaded task. The queueing delay multiplier both models and affects the queueing delay, and thus acts as an admission rule parameter. If the multiplier is too high, tasks are discouraged to offload, resulting in a lightly loaded MEC server which entails a low multiplier value. On the other hand, a too low value for the multiplier will result in more offloads, yielding high queueing delays. Therefore, there should be an optimal queueing delay multiplier value that balances the demand for the MEC server. We propose an adaptive scheme that finds an optimum value for the queueing delay multiplier on the fly, using damped averaging. We show that high priority users experience much lower average sojourn times compared to the low priority users. We also illustrate the effect of our method on the energy consumption of the mobile. Using a stand-alone simulator, we demonstrate the performance of the proposed method in several different scenarios with numerical experimentation.