Sommario
Embedded systems are increasingly defined by their networked and data-driven, computationally-intensive behavior. This includes Cyber-physical systems (CPS) and the Internet of Things (IoT), where edge computing has emerged as a way to provide scalability, latency and privacy by processing data near its source or sink. These are all examples of networks-of-systems (NoS) in which computations are distributed over clusters of resource-constrained edge devices that communicate over often open, public networks. Beyond traditional system-level design, this creates new challenges to tackle optimized mapping, offloading and scheduling of distributed applications while providing real-time, performance and other correctness guarantees. In this talk, we present our on-going work in this new era of network-level design and design automation. This first and foremost requires novel modeling approaches, including both formal models of computation and communication (MoCC) for network-level specification as well as NoS performance models and simulators. We will present a novel Reactive and Adaptive Dataflow (RADF) MoCC that addresses the problem of how to soundly express, analyze and realize deterministic real-time computation over unreliable and unpredictable networks. We will further discuss our work on fast yet accurate simulation of NoS and IoT systems. Such models can then drive automated synthesis and design space exploration of NoS applications and architectures, where we have developed associated compilers and runtime systems. We have applied such network-level design methods to specifically achieve an optimized, distributed deployment of deep learning applications on tightly constrained IoT edge clusters, which is enabled by novel partitioning and fusing methods and an adaptive work stealing middleware for dynamic workload distribution and load balancing. We will present our DeepThings framework and demonstrate results showing scalable CNN inference on a cluster of Raspberry Pi devices with less than 20MB of memory each.
Biografia
Andreas Gerstlauer is an Associate Professor in Electrical and Computer Engineering at The University of Texas at Austin. He received his Ph.D. in Information and Computer Science from the University of California, Irvine (UCI) in 2004. Prior to joining UT Austin in 2008, he was an Assistant Researcher in the Center for Embedded Computer Systems (CECS) at UC Irvine, leading a research group to develop electronic system-level (ESL) design tools. Commercial derivatives of such tools are in use at the Japanese Aerospace Exploration Agency (JAXA), NEC Toshiba Space Systems and others. Dr. Gerstlauer is co-author on 3 books and more than 100 publications. His work was recognized with the 2016 DAC Best Research Paper Award, the 2015 SAMOS Best Paper Award, and as one of the most most influential contributions in 10 years at DATE in 2008. He received a 2016-2017 Humboldt Research Fellowship and has presented in numerous industry and conference tutorials. He currently serves an Associate Editor for ACM TECS and TODAES journals, and he has served as Topic, Track or Program Chair of major international conferences such as DAC, DATE, ICCAD and CODES+ISSS. His research interests include system-level design automation, system modeling, design languages and methodologies, and embedded hardware and software synthesis.