GPU Acceleration in Unikernels Using Cricket GPU Virtualization
Niklas Eiling, Martin Kröning, Jonathan Klimt, Philipp Fensch, Stefan Lankes, Antonello Monti · 2023
Today, large compute clusters increasingly move towards heterogeneous architectures by employing accelerators, such as GPUs, to realize ever-increasing performance. To achieve maximum performance on these architectures, applications have to be tailored to the available hardware by using special APIs to interact with the hardware resources, such as the CUDA APIs for NVIDIA GPUs. Simultaneously, unikernels emerge as a solution for the increasing overhead introduced by the complexity of modern operating systems and their inability to optimize for specific application profiles. Unikernels allow for better static code checking and enable optimizations impossible with monolithic kernels, yielding more robust and faster programs. Despite this, there is a lack of support for using GPUs in unikernels. Due to the proprietary nature of the CUDA APIs, direct support for interacting with NVIDIA GPUs from unikernels is infeasible, resulting in applications requiring GPUs being unsuitable for deployment in unikernels.