CAREER: Operating System Support for Ephemeral and Malleable Accelerators

随着摩尔定律的衰落，计算机系统越来越依赖专门的硬件或加速器来提高性能、可扩展性和能源效率。预计到2025年，数据中心将消耗全球9%至20%的总电力，因此提高性能和效率对我们这个星球的健康至关重要。现场可编程门阵列，或称现场可编程门阵列，是一种可重新配置的硬件，可根据需要灵活地成为任何专用硬件。它们在现代数据中心很有吸引力，因为它们允许从单个可重新配置的设备合成各种加速器；亚马逊和微软等提供商现在支持云中的FPGA加速。然而，对于旨在管理固定功能设备的传统操作系统来说，向可重新配置硬件的范式转变是巨大的。当前系统假设单个应用程序独占访问硬件，而共享支持的缺乏限制了利用率和效率。本研究的目标是开发系统级抽象和端到端操作系统(OS)支持受保护的FPGA共享，从而将利用率和吞吐量提高一个数量级或更多。这是具有挑战性的，因为FPGA加速器是短暂的和可塑性的：它们可能会频繁变化，并能够随着FPGA资源份额的增加而进行扩展。这些特性与数十年来的操作系统设计假设背道而驰。该项目的目标是建立一个完整的开源堆栈，支持服务器进程高效、高利用率的保护性共享，将计算分流到现场可编程门阵列，并将其扩展为云FPGA平台的兼容层。对一个名为AmorOS的初始原型的测量表明，该方法很有希望，对于运行神经网络(DNN)推理的Microsoft Catapult和Amazon F1系统，该方法可以将利用率和吞吐量分别提高8倍和32倍。该系统有可能在支持现场可编程门阵列的数据中心广泛使用，工作结果可以指导下一代操作系统的设计，用于异构和可重构的硬件。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

With the decline of Moore's Law, computer systems increasingly rely on specialized hardware, or accelerators to improve performance, scalability and energy efficiency. Data centers are projected to consume between 9% and 20% of the world's total power by 2025, so improved performance and efficiency are important to the health of our planet. Field Programmable Gate Arrays, or FPGAs, are reconfigurable hardware with flexibility to become any specialized hardware on demand. They are compelling in modern data centers because they enable a wide range of accelerators to be synthesized from a single reconfigurable device; providers such as Amazon and Microsoft now support FPGA acceleration in the cloud. However, the paradigm shift to reconfigurable hardware is massive for traditional operating systems, which are designed to manage fixed-function devices. Current systems assume a single application has exclusive access to the hardware, and the lack of sharing support limits utilization and efficiency.The goal of this research is to develop system-level abstractions and end-to-end operating system (OS) support for protected FPGA sharing that improves utilization and throughput by an order of magnitude or more. This is challenging because FPGA accelerators are ephemeral and malleable: they may change frequently and be able to scale with increased shares of FPGA resource. These properties run counter to decades-old OS design assumptions. The project aims for a complete open-source stack that supports efficient high-utilization protected sharing for server processes to offload computations to FPGAs, and will extend it to function as a compatibility layer for cloud FPGA platforms. Measurements of an initial prototype called AmorphOS show that the approach has promise and can improve utilization and throughput by 8X and 32X for Microsoft Catapult and Amazon F1 systems running neural network (DNN) inference. The system has the possibility to be used pervasively in data centers supporting FPGAs, and the results of the work could guide the design of the next generation of operating systems for heterogeneous and reconfigurable hardware.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.