Towards Hardware as a Reconfigurable, Elastic, and Specialized Service

April 21, 2021

As modern Data Center workloads become increasingly complex, constrained, and critical, mainstream CPU-centric computing has had ever more difficulty in keeping pace. Future data centers are moving towards a more fluid and heterogeneous model, with computation and communication no longer localized to commodity CPUs and routers. Next generation data-centric Data Centers will compute everywhere, whether data is stationary (e.g. in memory) or on the move (e.g. in network). While deploying FPGAs in NICS, as co-processors, in the router, and in Bump-in-the-Wire configurations is a step towards implementing the data-centric model, it is only part of the overall solution. The other part is actually leveraging this reconfigurable hardware. For this to happen, two problems must be addressed: code generation and deployment generation. By code generation we mean transforming abstract representations of an algorithm into equivalent hardware. Deployment generation refers to the runtime support needed to facilitate the execution of this hardware on an FPGA. Efforts at creating supporting tools in these two areas have thus far provided limited benefits. This is because the efforts are limited in one or more of the following ways: They i) do not provide fundamental solutions to a number of challenges, which makes them useful only to a limited group of (mostly) hardware developers, ii) are constrained in their scope, or iii) are ad hoc, i.e., specific to a single usage context, FPGA vendor, or Data Center configuration. Moreover, efforts in these areas have largely been mutually exclusive, which results in incompatibility across development layers; this requires wrappers to be designed to make interfaces compatible. As a result there is significant complexity and effort required to code and deploy efficient custom hardware for FPGAs; effort that may be orders-of-magnitude greater than for analogous software environments. The goal of this dissertation is to create a framework that enables reconfigurable logic in Data Centers to be targeted with the same level of effort as for a single CPU core. The underlying mechanism to this is a framework, which we refer to as Hardware as a Reconfigurable, Elastic and Specialized Service, or HaaRNESS. In this dissertation, we address two of the core challenges of HaaRNESS: reducing the complexity of code generation by constraining High Level Synthesis (HLS) toolflows, and replacing ad hoc models of deployment generation by generalizing and formalizing what is needed for a hardware Operating System. These parts are unified by the back-end of HLS toolflows which link generated compute pipelines with the operating system, and provide appropriate APIs, wrappers, and software runtimes. The contributions of this dissertation are the following: i) an empirically guided set of systematic transformations for generating high quality HLS code; ii) a framework for instrumenting HLS compiler to identify and remove optimization blockers; iii) a framework for RTL simulation and IP generation of HLS kernels for rapid turnaround; and iv) a framework for generalization and formalization of hardware operating systems to address the ad hoc’ness of existing deployment generation and ensure uniform structure and APIs.