Applications and Programming Models for Large-Scale Heterogeneous Computing with FPGAs

随着我们进入云和物联网(IoT)时代，围绕计算能力和性能的问题变得越来越重要。使用微处理器的计算仍然是所有形式计算的中流砥柱，但微处理器不再足够，因为它们在功率或性能方面并不总是最有效的。因此，人们更加迫切地探索非传统计算方法。2015年，英特尔以167亿美元的价格收购了Altera，这让世界意识到了使用现场可编程门阵列(FGA)的好处，但也让计算世界变得更加多样化。对于许多应用，现场可编程门阵列既能提供性能优势，又能以更低的功耗需求做到这一点。问题是它们很难使用。我们工作的动机和挑战是让所有应用程序开发人员都可以访问现场可编程门阵列。这项建议侧重于探索在高性能计算中使用分布式现场可编程门阵列的编程模型和驱动应用程序，但许多技术也可以应用于物联网世界的计算频谱的另一端。*这项工作的长期目标是无缝地使用现场可编程门阵列，以便应用程序开发人员可以专注于他们试图解决的问题，而不是像今天那样努力让现场可编程门阵列工作。短期目标是为结合了现场可编程门阵列的不同系统建立编程基础设施，并用示范应用程序对其进行测试和验证。我们的总体方法是始终构建工作系统，因为这是证明我们的想法将会奏效的唯一方法，我们可以进行易于证明的测量。*虽然我们的目标是改善程序员使用现场可编程门阵列的体验，但该解决方案跨越整个计算平台或计算“堆栈”，将整个项目放在计算系统领域。现场可编程门阵列与微处理器有非常不同的特性，因为它们的编程和管理方式以及它们所具有的本机功能。这意味着它们不容易适应现有的基于微处理器的基础设施，我们整体工作的一个重要方面是找出如何适应它们，或者如何改变基础设施以更好地适应它们。*英特尔收购Altera颠覆了计算机界，并将可重构计算的兴趣从一个小社区扩大到更大的计算机体系结构社区，因为人们相信，现在使用FPGA可以带来显著的商业利益。由于英特尔，现场可编程门阵列不再是一种小众技术。拟议项目最重要的影响是解决可编程性问题，使应用程序开发人员能够使用现场可编程门阵列，从而满足性能和功率的限制。建议的工作适用于从物联网到高性能计算的所有计算领域。

As we enter into the era of the Cloud and the Internet of Things (IoT), issues around the power and performance of computing are becoming ever more important. Computing using microprocessors is still the mainstay of all forms of computing but microprocessors are no longer enough because they are not always the most efficient in terms of power or performance. As a result, non-traditional computing approaches are being explored with more urgency. The acquisition of Altera by Intel in 2015 for USD$16.7B has woken up the world to the benefits of using Field-Programmable Gate Arrays (FPGAs), but also made the computing world more heterogeneous. For many applications, FPGAs can provide both a performance benefit and do it with lower power requirements. The problem is that they are difficult to use. The motivation and challenge of our work is to make FPGAs accessible by all application developers. This proposal focuses on the exploration of programming models and driving applications for using distributed FPGAs in high-performance computing, but many of the techniques can also be applied at the other end of the computing spectrum in the IOT world.******The long-term objective of this work is to make the use of FPGAs seamless so that application developers can focus on the problem they are trying to solve, rather than struggling with making FPGAs work, as it is today. The short-term objectives are to build programming infrastructures for heterogeneous systems incorporating FPGAs and test and validate them with exemplary applications. Our overall methodology is to always build working systems because that is the only way to prove that our ideas will work, and we can make measurements that are easy to justify.******While our goal is to improve the programmer experience of using FPGAs, the solution spans the entire computing platform, or computing "stack", putting the overall project in the area of computing systems. FPGAs have very different properties than microprocessors because of the way that they are programmed and managed, and the native capabilities that they have. This means that they do not easily fit into the existing microprocessor-based infrastructure and an important aspect of our overall work is to figure out how to fit them, or how to change the infrastructure to accommodate them well.******The Intel acquisition of Altera has disrupted the computing world and broadened the interest of reconfigurable computing from a small community to the much larger computer architecture community because of the belief that there are now significant commercial benefits from using FPGAs. FPGAs are no longer a niche technology because of Intel. The most significant impact of the proposed project is to address the programmability issue and enable application developers to use FPGAs so that the constraints of performance and power can be met. The proposed work is applicable across the computing spectrum, from the IOT to high-performance computing.***********