CNS Core:Small: Toward Automatic Virtualization of Accelerators

点击翻译按钮获取中文摘要

The cloud has transformed computing and society over the past decade by providing fast on-demand access to scalable computing resources. At the same time, the end of Moore’s law has driven the emergence of a panoply of hardware accelerators such as GPUs, TPUs, and FPGAs that are critical to game-changing workloads such as machine learning. Unfortunately, cloud migration and hardware specialization are on a collision course: cloud applications run on virtual infrastructure, but practical virtualization techniques for accelerators remains an open and challenging problem. Solutions offered by cloud providers sacrifice the consolidation benefits of virtualization and waste energy. With data centers projected to consume up to 20% of the world's total power by 2025, improving performance and efficiency of cloud infrastructure is becoming critical to the health of our planet.The proposed research will address key obstacles that limit efficiency and wide adoption of hardware acceleration in the cloud, focusing on two questions. The first is how to build practical, efficient virtualization support for emerging hardware-accelerators. The second is how to make accelerators of the future more efficiently virtualizable. The PI will develop language, OS, and runtime techniques to automate the construction of virtual accelerator support. Applying these techniques to cloud accelerators is a significant research challenge. The PI will explore techniques to dramatically improve virtualization support provided directly by accelerators and will develop language-level virtualization support for reconfigurable accelerators such as field programmable gate arrays. By bringing efficient virtualization support to cloud accelerators and cloud services that use accelerators, the proposed research will bring power-efficient cloud computing to new problems and domains, and reduce energy waste.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.