Operating Systems Must Support GPU Abstractions.

This paper argues that lack of OS support for GPU abstractions fundamentally limits the usability of GPUs in many application domains. OSes offer abstractions for most common resources such as CPUs, input devices, and file systems. In contrast, OSes currently hide GPUs behind an awkward ioctl interface, shifting the burden for abstractions onto user libraries and run-times. Consequently, OSes cannot provide system-wide guarantees such as fairness and isolation for GPUs, and developers must sacrifice modularity and performance when composing systems that integrate GPUs along with other OS-managed resources. We propose new kernel abstractions to support GPUs and other accelerator devices as first class computing resources.