CSR: Small: Redshift: An Operating System for Pervasive Hardware Acceleration

University Of Utah, Salt Lake City UT

Investigators

Abstract

In contrast to today's systems centered around general-purpose processors also known as central processing units (CPUs), the next generation of high-performance computers will inherently rely on diverse, heterogeneous hardware ranging from many-core processors like Intel Xeon Phi that contains up to 72 processor cores and graphical processing units (GPUs) to specialized hardware accelerators, like specialized machine-learning chips and field-programmable gate arrays (FPGAs) re-programmed on demand for a specific task. In a hardware-accelerated environment that consists of many diverse execution units, the execution of a program is no longer a conventional thread tied to a single CPU, but a graph of small computations scheduled on a set of hardware accelerators each implementing a part of the program logic. Redshift is a new operating system for developing applications that leverage performance of a heterogeneous hardware-accelerated system. At the core of Redshift is a dataflow programming model that enables execution of commodity programs on a network of heterogeneous hardware execution units with only minimal modifications. Redshift implements programs as collections of asynchronous invocations that transparently move execution between hardware functions. A novel runtime maps computations to execution units, balances load among them, and scales the hardware graph of computation in response to load. The problem of efficient computing environments has large impacts on society as a whole: a rapidly growing share of scientific discoveries is done in silico. As this trend continues, we as a society depend on computational capacity of modern computers. Redshift will provide a foundation for developing the next generation of computation intensive applications in the areas of artificial intelligence, e.g., speech and image recognition, personal digital assistance, big-data analytical applications, genomic and personalized medicine, drug discovery, and many more. Redshift will be implemented as a practical layer compatible with de facto research and industry standard Linux operating system, and will be open source, directly benefiting the broader community. To make our approach widely available, Redshift will be hosted in a publicly-readable repository, and will be available to anyone (https://github.com/mars-research/redshift). Additionally, as parts of Redshift will be developed in the openly-available National Science Foundation-funded CloudLab and Emulab testbeds, Redshift will be available for a test drive via a CloudLab profile (https://www.cloudlab.us/p/redshift/testdrive) that automatically instantiates a collection of nodes running Redshift). 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.