GTC Silicon Valley-2019: Dynamic Sharing of GPUs and IO in a PCIe Network
GTC Silicon Valley-2019 ID:S9709:Dynamic Sharing of GPUs and IO in a PCIe Network
Haakon Stensland(Simula Research Laboratory)
Learn how to efficiently share GPUs, NVMe drives, and other IO devices in a PCI Express network using Dolphin Interconnect Solutions' SmartIO. We'll explain how we eliminated restrictions of traditional IO devices, which are statically assigned to a single root complex (host machine) and lack flexible support for features such as hot-add, device migration, and remote access without complex software frameworks. We'll highlight how SmartIO provides a framework that allows PCIe devices and systems to use remote systems to flexibly access devices such as GPUs, NVMe drives, and other IO devices. We will demonstrate how we implemented SmartIO using standard PCIe and non-transparent bridging and show how our system got near-native performance when moving data borrowed GPUs and NVMe drives. We'll also explore how we can dynamically add more GPUs to scale performance.