Design and Evaluation of A Prediction-based Dynamic Edge Computing System

We investigate a mobile edge computing environment where edge computing nodes provide their computation capacities to process the computation intensive tasks submitted by end users. We introduce a Cloudlet Assisted Cooperative Task Assignment (CACTA) system that organizes edge nodes that are geographically close to a user into a cluster to collaboratively work on the user's tasks. The system enables a user to minimize his/her total cost which is a weighted combination of latency (i.e., the task's completion time), and the costs incurred in working on the task. The total cost captures the tradeoff that the user would like to make between latency and computing related costs. It is challenging for the system to find an optimal strategy that assigns workload to edge nodes to meet the user's optimization goal, due to the time-varying available capacities and the mobility of edge nodes. To address the challenge, we model the system as a discrete time system in which each edge node's capacity and cost vary over different time slots, and the system assigns parts of the task to the edge nodes in the cluster over time. We introduce a prediction-based dynamic task assignment algorithm, referred to PA-OPT, that assigns workload to edge nodes in each time slot based on the prediction of their capacities/costs and an empirical optimal allocation strategy which is learned from an offline optimal solution from historical data. Then we apply our system design to a video data analysis application, and conduct extensive simulations driven by a Google cloud data trace. We have demonstrated that our proposed algorithm/system achieves significantly higher performance than several other algorithms, and especially its performance is very close to that of an offline optimal solution.