Boosting GPGPU virtualization and multiplexing with RDMA communication

Hardware virtualization is essential in High-Performance Computing (HPC) and Cloud Computing. It addresses resource allocation challenges, scalability, and performance isolation in shared environments. With the rise of GPGPU-accelerated HPC clusters, research has increasingly focused on GPGPU virtualization, leading to various solutions. This work explores the Generalized Virtualization Service (GVirtuS), a transparent virtualization solution with a plug-in framework for easy development and choice of communicators and stub libraries. We discuss GVirtuS’ implementation and design choices and compare them with similar solutions. Additionally, we show how GVirtuS’ performance was enhanced with a novel RDMA-based communicator, evaluated through CUDA implementations of SAXPY and Matrix Multiplication. Our RDMA Communicator achieved a 35% performance boost over the TCP Communicator on Infiniband and a 55% boost over TCP on Ethernet.