Composability of Cloud Accelerators in Virtual World Simulations

Abstract

With the rise of computer vision as a driving force in today's digitized world, virtual world simulations (VWS) have become a critical application class, requiring high-performance and versatile computing systems. Immersive and interactive experiences offered by VWS are essential in domains such as gaming, engineering, and education. Cloud computing provides a compelling solution for VWS, with its scalability, cost-effectiveness, and accessibility, combined with efficient processing capabilities of hardware accelerators such as FPGAs and GPUs. This paper presents APEIRON, a distributed library built on the Ray framework, which enables easy composition of multiple hardware accelerators on popular public clouds, like AWS, and emerging disaggregated systems like IBM cloudFPGA. The design and implementation of APEIRON is outlined in this paper, providing a developer-friendly solution for meeting the growing demands of computer vision in VWS. Experiments demonstrate APEIRON's competitive performance for deploying VWS on the cloud, compared to existing libraries.