Modern smartphones serve as an integral component in emerging distributed applications, including those spanning the domains of crowdsensing, smart homes, and mobile health. A unique aspect of these applications is their transition towards using machine learning-based techniques that provide state-of-the-art performance. We show that for distributed applications using machine learning, timing errors across smartphones can severely degrade application performance. We highlight the challenges to achieve precise time synchronization across smartphones due to their limited hardware support, exclusively wireless networking interface, and restricted timing stack control which results in timing errors in the order of multiple seconds across Android smartphones. As a solution, we present alternative approaches for time synchronization by exploiting the various peripherals available on modern smartphone devices. Our evaluation across Android smartphones typically achieves synchronization accuracy within (i) 200 microseconds using audio, (ii) 3000 microseconds using Bluetooth Low Energy, and (iii) 1000 microseconds using Wi-Fi. The provided guide and accompanying open-source implementations offer developers a means to select the appropriate time synchronization technique when building distributed applications across smartphone devices.

Authors: Sandeep Singh Sandha and Mani Srivastava (mbs@ucla.edu)
University of California, Los Angeles.