Efficient and Scalable Computing for Resource-Constrained Cyber-Physical Systems: A Layered Approach
Date: 2021/04/06 - 2021/04/06
Academic Seminar: Efficient and Scalable Computing for Resource-Constrained Cyber-Physical Systems: A Layered Approach
Speaker: Dr. An Zou, Washington University in St. Louis
Time: 9:00a.m.-10:00a.m., April 6th, 2021 (Beijing Time)
Location: via Zoom
With the evolution of computing and communication technology, cyber-physical systems such as self-driving cars, unmanned aerial vehicles, and mobile cognitive robots are achieving increasing levels of multifunctionality and miniaturization, enabling them to execute versatile tasks in a resource-constrained environment. Therefore, the computing systems that power these resource-constrained cyber-physical systems (RCCPSs) have to achieve high efficiency and scalability. First of all, given a fixed amount of onboard energy, these computing systems should not only be power-efficient but also exhibit sufficiently high performance to gracefully handle complex algorithms for learning-based perception and AI-driven decision-making. Meanwhile, scalability requires that the current computing system and its components can be extended both horizontally, with more resources, and vertically, with emerging advanced technology. To achieve efficient and scalable computing systems in RCCPSs, my research broadly investigates a set of techniques and solutions via a bottom-up layered approach. This layered approach leverages the characteristics of each system layer (e.g., the circuit, architecture, and operating system layers) and their interactions to discover and explore the optimal system tradeoffs among performance, efficiency, and scalability. At the circuit layer, we investigate the benefits of novel power delivery and management schemes enabled by integrated voltage regulators (IVRs). Then, between the circuit and microarchitecture/architecture layers, we present a voltage-stacked power delivery system that offers best-in-class power delivery efficiency for many-core systems. After this, using Graphics Processing Units (GPUs) as a case study, we develop a real-time resource scheduling framework at the architecture and operating system layers for heterogeneous computing platforms with guaranteed task deadlines. Finally, fast dynamic voltage and frequency scaling (DVFS) based power management across the circuit, architecture, and operating system layers is studied through a learning-based hierarchical power management strategy for multi-/many-core systems. On the one hand, each part either opens new opportunities at other layers or leverages co-designs across several layers. On the other hand, the four parts work together to form a complete, layered approach, spanning from the circuit layer, through the architecture layer, to the operating system and application layers. This layer-spanning approach successfully improves both power and performance efficiencies of the computing systems, which further strengthens whole cyber-physical systems.
An Zou received his B.S., M.S degrees from Harbin Institute of Technology in 2013 and 2015, and Ph.D. in Electrical Engineering from Washington University in St. Louis in 2021. His research focuses on computer architecture and embedded systems. He broadly investigated a set of techniques and solutions via a bottom-up layered approach to improve computing power and performance efficiency. His work has been extensively published and recognized at top-tier conferences and journals. His work received the best paper nominations at DAC (2017) and MLCAD (2020). He is the recipient of the A. Richard Newton Young Student Fellowship (2017) and a Chinese National Scholarship (2014).