$\pi$-Configurator: Enabling Efficient Configuration of Pipelined Applications on the Edge

Abstract

Modern edge computing applications involve a computational pipeline of multiple stages. Each stage typically involves many configuration options that affect application-level quality of service. Identifying an optimal configuration is challenging but important when the applications run under resource constraints. The main challenge is that when pipelines have many stages and each stage has many settings, the overall configuration state space is exceedingly large. We propose $\pi$-Configurator, a system for sampling application-level quality of service (QoS) metrics, constructing an approximation of the configuration state space, and finally identifying an optimal configuration for the application. We demonstrate the accuracy and effectiveness of $\pi$ - Configurator with four multi-stage data processing applications on resource-limited edge computing platforms. $\pi$-Configurator incurs low approximation error, and is one to two orders of magnitude faster than complete sampling approaches. The configurations identified by $\pi$-Configurator outperform those identified by existing local adaptation approaches by 99%.