Paper Title: Multi-Host Sharing of NVMe Drives and GPUs Using PCIe Fabrics

Paper Abstract: As the use of GPUs for deep learning, artificial intelligence and machine learning increases, so does the demand for a more efficient deployment of GPU and NVMe resources. This topic will discuss how PCIe fabric switches can be used to create a flexible, low-latency, high-performance fabric interconnect to a shared pool of GPUs and NVMe SSDs while still supporting standard host operating system drivers. Using dynamic partitioning and multi-host SR-IOV sharing techniques, GPU and NVMe resources can be ‘composed’ or dynamically allocated to a specific host or set of hosts allowing real-time allocation of resources to match workload requirements. Key concepts of PCIe fabrics and multi-host sharing of SR-IOV devices will also be covered.

Paper Author: Vincent Hache, Principal Applications Engineer, Microsemi

Author Bio: Vincent is the Firmware Principal Applications Engineer for Microsemi's Switchtec PCIe Switches. He is also the software architect and product owner of ChipLink, Microsemi's advanced cross-platform debug and diagnostics tool. His current focus is on enabling PCIe fabrics using advanced fabric switches to provide a flexible interconnect for shared pools of GPUs or NVMe SSDs. The main purpose is to accelerate AI and machine learning applications. Vincent has been providing architectural and engineering support for Tier 1 customers storage, server, and data center designs in NVMe and GPU interconnect, and Serial Attached SCSI (SAS) and PCIe infrastructure for the last 6 years. He earned a Bachelor of Electrical Engineering degree from Carleton University (Canada).