Edge replication strategies for wide-area distributed processing

Abstract

The rapid digitalization across industries comes with many challenges. One key problem is how the ever-growing and volatile data generated at distributed locations can be efficiently processed to inform decision making and improve products. Unfortunately, wide-area network capacity cannot cope with the growth of the data at the network edges. Thus, it is imperative to decide which data should be processed in-situ at the edge and which should be transferred and analyzed in data centers. In this paper, we study two families of proactive online data replication strategies, namely ski-rental and machine learning algorithms, to decide which data is processed at the edge, close to where it is generated, and which is transferred to a data center. Our analysis using real query traces from a Global 2000 company shows that such online replication strategies can significantly reduce data transfer volume in many cases up to 50% compared to naive approaches and achieve close to optimal performance. After analyzing their shortcomings for ease of use and performance, we propose a hybrid strategy that combines the advantages of both competitive and machine learning algorithms.