Joint Job Offloading and Resource Allocation for Distributed Deep Learning in Edge Computing

Abstract

Under the paradigm of Edge Computing, the enormous data generated at the network edge can be processed locally. Machine learning methods are often adopted to make full utilization of these data, among which the deep learning is a promising one. When considering the inherent distributed feature of these data, it is appeal to conduct distributed deep learning tasks directly at the edge. We focus on an edge computing system that conduct distributed deep learning tasks using gradient descent based approaches. To ensure the system's performance, there are at least two major challenges to cope with: how to offload the training jobs with multiple data source nodes and how to allocate the limited resources on each edge server among training jobs. In this paper, we jointly consider the two challenges, aiming to maximize the system throughput while ensuring the system's quality of service (QoS). We formulate the joint problem as an integer non-linear program and propose an efficient approximation algorithm based on reformulation and randomized rounding technique. Simulation results prove that the proposed algorithm can improve 56% of the system throughput and 53% of resource utilization when compared to the conventional baseline algorithms.