Chong Song of KAIST Graduate School of AI and School of Electrical Engineering Kim Se-yeon Ph .D student and Professor Lee Kyeong-han of Seoul National University Department of Electrical and Computer Engineering received the ACM MobiSys Best Paper Award which were held online due to COVID-19 situation. Lee studied his Ph. D and Kim is studying in Chongs lab.

ACM MobiSys is the top academic conference in mobile systems with a total of 266 papers submitted this year and 36 of them accepted. As the lead author being Asian Chong’s team is the first to receive the award. The paper was published under the name of ‘zTT: Learning-based DVFS with Zero Thermal Throttling for Mobile Devices’.

The paper is mainly about the Dynamic Voltage and Frequency Scaling (DVFS) which refers to the technology that manipulates voltage and frequency to secure processor performance and minimizing energy consumption in an OS level. There are two main obstacles that the research targets to overcome.

First the performance of individual applications is not secured since it is operated in OS level. Second it may harm the quality of experience by causing thermal throttling because it may not sense the frequent transition of mobile environment. Especially overheat in mobile devices is a stubborn problem. The research suggests a Deep-Reinforcement Learning based frequency distributing technology that prevents overheating and lowering the energy consumption. It catches the frequent change and adapts the system instantaneously resulting in lower energy consumption and stable performance. The team expects the tech to package the OS and applications into a single set-up option to result an optimal performance.

Kim explained that the research is about alleviating thermal throttling and following performance decline by Deep-Reinforcement Learning based dynamic voltage/frequency scaling.

Professor Chong said that energy distribution between processor components such as CPU and GPU was important in order to realize improvement in performance and prevention of thermal throttling. Conventional methods could not encompass the change of physical environment such as temperature and cooling condition and the features of individual applications. Mobile platform was one of the last fields that the AI methods to be applicated but the research team solved energy problem and showed that AI and reinforced learning could contribute. It somehow showed potential of AI and machine learning technology and opened possibility to be applicated in future OS.