CAREER: An adaptive framework to accelerate real-time workloads in heterogeneous and reconfigurable environments

点击翻译按钮获取中文摘要

Artificial intelligence and machine learning are enabling real-time decisions based on live data for interactive scientific discovery and mission critical applications such as autonomous driving and smart grid. They are increasingly powered by heterogeneous and even reconfigurable accelerators. The reconfigurability and heterogeneity of accelerators, together with stringent performance requirements and complex dependencies in real-time workloads, bring daunting operational challenges. These issues, if left unaddressed, would slow down scientific discovery and waste lots of computing resources and energy. This project will develop a heterogeneity and reconfigurability aware framework to accelerate real-time artificial intelligence and machine learning without hurting other workloads. It will benefit the society by improving the efficiency of costly computing systems, which saves taxpayers' money and better utilize existing investments. Real-time artificial intelligence and machine learning powered by the framework can better serve the society, e.g., accelerating scientific discovery and enabling data-driven control. The project will bring innovative education, outreach and training opportunities for both academic and industrial participants to train the next generation of researchers and practitioners for the society.Today, managing heterogeneous and reconfigurable systems for diverse workloads with high resource utilization and performance guarantee is an extremely challenging task. This project will design and implement an adaptive framework which automatically detects, profiles, and analyzes both workloads and accelerators on the fly. Based on the information, it adaptively reconfigures them to match resource capabilities with workload needs. Global and local optimization will be used to accommodate multiple types of workloads and the configuring, partitioning, placement, scheduling, and execution of models in each workload. The developed framework will provide provable performance even with partial information in unknown environments, which is urgently needed due to the ever increasing system complexity and volatility in workloads. Novel global resource allocation policies will be developed based on optimization techniques in this project to provide performance guarantee such as fairness, strategyproofness, and Pareto efficiency. Throughout the project, a reciprocal methodology is envisioned: the framework accelerates artificial intelligence/machine learning workloads and artificial intelligence/machine learning techniques enable the framework.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.