LiveDewStream: A stream processing platform for running in-lab distributed deep learning inferences on smartphone clusters at the edge

Abstract

Dew computing, an evolution of Fog computing, aims at fulfilling computing needs, such as deep learning applied to object classification, close to where data is originated and using computing resources that include consumer electronic devices such as smartphones. Simulation tools like DewSim aid the study of resource allocation mechanisms for exploiting clusters of smartphones, however, there is a gap w.r.t software tools that allow to perform similar studies over real Dew computing testbeds. We have developed LiveDewStream, an open source project to model executable tasks derived from data streams to be run on real smartphone clusters. The project offers a key functionality missing in other tools: reproducibility of battery-driven Dew experiments. Our major contribution is to provide the community a common in vivo platform to study best-performing allocation mechanisms under different stream processing scenarios and/or deep learning inference models.