Towards AI-Driven Mobile Systems
Sangtae Ha is an Associate Professor in the Department of Computer Science at the University of Colorado Boulder. He received his Ph.D. in Computer Science from North Carolina State University and was an Associate Research Scholar at Princeton University. He leads the Internet Systems Lab (ISL), where more than a dozen researchers, including Ph.D. students and postdocs, conduct computer systems research, including machine learning and deep learning systems, networks and distributed systems, mobile systems, internet protocols and algorithms, video streaming, storage systems, and wireless networks and security.
His research has been sponsored by many funding agencies and industries, including NSF, DARPA, NIST, AT&T Research, Cisco Systems. Previously he open-sourced CUBIC, the default TCP congestion control algorithm on the Internet serving the majority of Internet servers worldwide and several billion Android and Linux devices. As a founding member, he currently serves the Chief Architect Officer at Earable (http://earable.ai), an artificial intelligence (AI)-based wearable startup. He is also a co-founder/advisor for three companies: Datami (http://www.datami.com), a startup on wireless networks, serving over 5 million daily active users in 45+ mobile operators, Zoomi (http://www.zoomi.ai), an artificial intelligence learning company and Myota (http://myota.io), a storage startup spun off from his recent research in secure and reliable storage.
He received the Best Paper Awards from ACM MobiSys in 2019 and 2021, AT&T Faculty Research Award and Samsung Global Research Outreach Award in 2017, and the INFORMS ISS Design Science Award in 2014.
In this talk, I will provide an overview of my group’s research and present my recent research on AI-driven mobile systems: PERCEIVE and zTT.
Our earlier work successfully infers the load of the cellular downlink locally from the client-side using machine learning (ML) to develop fully distributed transmissions among mobile devices that jointly optimize the spectral efficiency of cellular networks and battery consumption of those devices. PERCEIVE aims to solve the remaining story: estimating a load of cellular uplink locally from the client-side. It is a deep learning-based uplink throughput prediction framework exploiting a two-stage LSTM (Long Short-Term Memory) design. Our implementation of WebRTC, a de-facto video streaming platform from Google, integrated with PERCEIVE, shows that it could improve the video quality by up to 4x over the default WebRTC.
This machine learning could enhance legacy control and scheduling algorithms in computer systems? zTT aims to answer this question. DVFS (dynamic voltage and frequency scaling) is a system-level technique that adjusts voltage and frequency levels of CPU/GPU at runtime to balance energy efficiency and high performance. While DVFS has been studied for many years, it is challenging to optimize DVFS for mobile devices due to their unique characteristics – application diversity and temperature changes due to mobility. zTT is a deep reinforcement learning-based DVFS for mobile devices. zTT learns thermal environmental characteristics and jointly scales CPU and GPU frequencies to maximize the application performance in an energy-efficient manner while achieving zero thermal throttling. The performance evaluation shows that zTT successfully manages a rendering application in a high-temperature environment while saving 23.9% less power consumption while the default mobile DVFS fails to render at a target frame rate.