Spatiotemporal Analysis on Distributed Task Offloading in Extreme Edge Devices

Abstract

With the rapid development of the Internet of Things (IoT), the number of smart devices connected to the Internet is exponentially increasing, resulting in large-scale data and inadequate resources, which has caused high congestion and slow response delay in legacy cloud computing models. Multi-access Edge Computing (MEC) is a computing paradigm that can facilitate both delay-sensitive, and data-intensive tasks associated with IoT applications. MEC provides low latency by pushing the resources closer to the applications. However, due to the increase in the number of devices that use MEC, the high congestion problem remains unsolved. A promising solution is to take advantage of the abundant and underutilized computing resources of the Extreme Edge Devices (EEDs). EEDs bring the computing service closer to the end-users, which could significantly reduce the delay caused by cloud execution. However, the success of such an extreme edge parallel computing paradigm is impacted by i) wireless device-to-device (D2D)communication performance, a requirement for the communication between the recruited EEDs and the task requester to perform the offloading process, ii) the computing capabilities of the EEDs, which governs the execution time of each offloaded task and iii) the reliability of the recruited EEDs. In this context, a novel spatiotemporal framework employing stochastic geometry and absorbing continuous time Markov chains (ACTMC) is developed to analyze the communication and computation performance of extreme edge computing systems. Using this framework, we study the influence of various system parameters on the average task response delay over a baseline system, where the devices are recruited randomly and do not fail during the execution. Extensive evaluations have shown that the EED-enabled system outperforms MEC in terms of the average response delay in some cases. Next, we developed an advanced model, where the possibility of failure of the recruited EEDs is considered, and the impact of the recruitment criteria of EEDs on the recruitment time is investigated. Our findings have revealed the optimal number of slices that the task should be divided into to minimize the total computation time, which will minimize the average response delay, and how that optimal number is affected by the various system parameters.