Optimizing the Performance-Cost Tradeoff in Cross-Edge Analytics

Abstract

As an emerging computing paradigm, edge computing has gathered much attention recently. With edge computing, application datasets as examplified by performance logs and activity records naturally span across multiple edge sites. Analyzing such cross-edge datasets is however, by no means trivial, since the status quo approach which aggregates the raw data to a centralized cloud datacenter incurrs high performance and cost overheads. To address this challenge, in this paper, we tackle the problem of speeding up cross-edge analytics with low traffic cost, through joint optimization of task and input data placement. The resulted performance-cost tradeoff problem is difficult due its non-convexity and the uncertainty of query characteristics. To address these challenges, we combine convex relaxation with a two-stage optimization. Specifically, a prediction of the query characteristics is used to determine the data movement when the data is generated, and then the actual value is used to decide the task placement when the query arrives. Evaluations using a production trace from a Facebook cluster highlight that, the two-stage joint optimization approach can reduce the total cost by up to 83% compared to the status quo approach.