Enabling predictable parallelism in single-GPU systems with persistent CUDA threads

TL;DR

This paper introduces a method to achieve predictable parallelism in single-GPU systems by leveraging persistent CUDA threads, addressing the challenge of providing timing guarantees in real-time applications.

Contribution

It presents a novel approach that opens the GPU architecture for intra-GPU predictable execution, moving beyond treating the GPU as a monolithic device.

Findings

Enables predictable execution within modern GPU architectures.

Provides a framework for real-time guarantees in GPU computing.

Improves timing predictability for safety-critical applications.

Abstract

Graphics Processing Unit, or GPUs, have been successfully adopted both for graphic computation in 3D applications, and for general purpose application (GP-GPUs), thank to their tremendous performance-per-watt. Recently, there is a big interest in adopting them also within automotive and avionic industrial settings, imposing for the first time real-time constraints on the design of such devices. Unfortunately, it is extremely hard to extract timing guarantees from modern GPU designs, and current approaches rely on a model where the GPU is treated as a unique monolithic execution device. Unlike state-of-the-art of research, we try to open the box of modern GPU architectures, providing a clean way to exploit intra-GPU predictable execution.