Architecture and Automation Techniques for Resilient Computer Systems

点击翻译按钮获取中文摘要

Mitigating computer system failure has been a challenge since the dawn of modern computing, but a number of factors are now converging such that resilience (a system's capacity to operate in the presence of failure due to manufacturing defect or faults in the field) is now a first-class design constraint. Semiconductor scaling has made it possible to introduce computers into every aspect of our lives. Scaling, which makes the wires and transistors of integrated circuits increasingly miniature, unfortunately also makes them increasingly vulnerable to myriad sources of failure: small, low-power, high-performance devices tend to be defect-prone, susceptible to transient error, and vulnerable to performance degradation as well as early total breakdown. If resilience is not addressed, the resulting costs will limit the introduction of new products and jeopardize the steady growth on which the information technology economy depends. I therefore propose a research program focusing on the development of novel architectures and automation techniques needed to support the efficient, effective, design of resilient computer systems. My program will focus on three short-term objectives: (1) developing new, resilience-aware hardware/software (HW/SW) partitioning strategies; (2) developing resilient system-level architectures and organizations, specifically targeting (a) lifetime and yield improvement and (b) low-cost safety-critical systems; and, (3) developing parallel automation techniques which co-optimize systems at these various levels of design abstraction. This research program will lay a firm foundation of general principles of resilient computing for future research and development efforts targeting specific application domains, and additional opportunity for resilience co-optimization in the microarchitectural and circuit-level design of computer systems. Furthermore, it will give supported student research assistants the opportunity to develop important skills in parallel programming, especially using platforms such as OpenMP, OpenCL, and CUDA, all of which are increasingly applied in a growing number of important fields.