Learning scheduling bursty requests in Mobile Edge Computing using DeepLoad

Abstract

The emergence of Mobile Edge Computing (MEC) alleviates the large transmission latency resulting from the traditional cloud computing. For the compute-intensive requests such as video analysis, mobile users prefer to obtain a desired quality of experience (QoE) with neglected latency and reduced energy consumption. The popularity of smart devices allows users to release a run of compute-intensive as well as latency-sensitive requests anywhere, which may lead to bursty requests. A single resource-constrained edge server nearby is capable of handling a small amount of requests quickly, yet it seems helpless when encountering bursty compute-intensive requests. Despite the abundance of recently proposed schemes, the majority focus on efficiently scheduling pending requests in a single edge server, and ignored the potential role of edge collaboration to schedule bursty requests. Besides, while some recent studies proposed to finish a task using multiple devices, they focused on collaboration between mobile devices rather than between edge servers. Hence, we propose DeepLoad, a S2S system that schedules the bursty requests with a collaborative method using reinforcement learning (RL). DeepLoad decouples the scheduling decision into AP selection for setting the access point and workload redistribution for collaborative servers. DeepLoad trains a neural network model that picks decisions for each request based on observations collected by mobile devices. DeepLoad learns to make scheduling decisions solely through the resulting performance of historical decisions rather than rely on pre-programmed models or specific assumptions for the environment. Naturally, DeepLoad automatically learns the scheduling algorithm for each request and obtains a gratifying QoE. We aim to maximize the fraction of requests finished before their attached deadlines. Based on the Shanghai taxi trajectory data set, we design a simulator to obtain abundant samples, and leverage two GeForce GTX TITAN Xp GPUs to train the Actor–Critic network. Compared to the state-of-the-art bandwidth-based and server resources-based methods, DeepLoad can achieve a significant improvement in average fraction.