Operating and RuntimeSystems Towards an Efficient and Secure Edge
Abstract
Driven by emergingworkloads (e.g., on-device machine learning and sensor dataharvesting/processing), an edge platform adopts abundant and heterogeneoushardware, e.g. DSP/GPU/NPU for efficiency and TEE (e.g., Arm TrustZone) forsecurity and privacy. Yet, existing system software fails to harness the richhardware resources to foster the intended efficiency, security, as well asprivacy. On one hand, the hardware is poorly utilized by OS kernel anduserspace runtime, leading to energy waste. On the other hand, withoutsufficient system services, the near-baremetal TEE kernel inevitably relies onan untrusted OS kernel, creating security and privacy loopholes.
To address the abovechallenges, this dissertation presents our understanding of system designstowards an efficient and secure heterogeneous edge. Our philosophy is toco-design system software with hardware, which instantiates two key systemsdesign visions: 1) system software needs to specialize for workloads andhardware, and 2) it needs to proactively orchestrate hardware resources forexploiting their heterogeneity.
To demonstrate theefficacy of our ideas, we have built five systems running atop three diverseenvironments, spanning the mature Linux kernel, a TEE kernel, and a userspaceruntime. Through evaluation on a suite of commodity SoCs using comprehensivebenchmarks, we show that the five systems can work in harmony to improve theenergy and execution efficiency of an edge platform and mitigate its securityrisks as well as privacy leaks.
Moreover, the systemspresented in this dissertation enable many possibilities for the software runningon the edge which were considered impossible; they open new doors to futureresearch in the field.
Committee:
- YangfengJi, Committee Chair, (CS/SEAS/UVA)
- FelixLin, Advisor, (CS/SEAS/UVA)
- DavidEvans (CS/SEAS/UVA)
- MirceaStan (ECE/SEAS/UVA)
- Y.Charlie Hu (ECE/Purdue University)