Online Training Flow Scheduling for Geo-Distributed Machine Learning Jobs Over Heterogeneous and Dynamic Networks

Geo-Distributed Machine Leaning (Geo-DML) has been a promising technology, which performs collaborative learning across geographically dispersed data centers (DCs) with privacy-preserving over Wide Area Networks (WANs). Unfortunately, the limited and heterogeneous WAN bandwidth poses significant challenges to the performance of Geo-DML systems, leading to increased communication overhead and affecting the revenue of ISPs eventually. In particular, when multiple online jobs coexist in Geo-DML systems, the competition for bandwidth between training flows of different jobs aggravates this negative impact. To alleviate it, this paper investigates the problem of online training flow scheduling for Geo-DML jobs. We first formulate the studied problem as an Linear Programming (LP) model with the objective of maximizing the revenue of ISPs. Then, we propose an online traffic scheduling algorithm called Training Flow Adaptive Steering (TFAS), which exploits a primal-dual framework, tailored for efficient resource allocation of jobs to schedule training flows, such that system resources are maximally utilized and training procedures can be expedited and completed in a timely manner. Meanwhile, we conduct rigorous theoretical analysis to guarantee that the proposed algorithm can achieve a good competitive ratio. Extensive evaluation results demonstrate that our algorithm performs well and outperforms commonly adopted solutions 36.2%-49.4% in average.

I. Introduction

The advent of the Internet of Things (IoT) and the widespread adoption of 5G have resulted in an exponential increase in the amount of data generated by user devices. To accommodate the surge in data, Internet service providers (ISPs) have established many data centers (DCs) worldwide for the collection and storage of large datasets from users [1], [2]. Such vast data reservoir provides ample impetus for the large-scale implementation of machine learning (ML) systems to support AI-driven applications. The conventional method for training an ML model involves aggregating the extensive data scattered across different DCs into a single designated DC and subsequently training the ML model in a centralized manner [3]. However, this approach suffers from critical disadvantages such as high communication latency and heavy traffic loading. Additionally, concerns surrounding data security and privacy restrictions continue to impede the progress of the centralized training method [4]. To address these challenges, a novel ML paradigm, Geo-Distributed Machine Learning (Geo-DML), has been proposed. Geo-DML spans multiple DCs to enable privacy-preserving multi-party cooperative learning through Wide Area Networks (WANs) [5].