NVMe over thin Kernel using DPDK and user-mode file system

0 0
  • 0 Collaborators

Implement NVMe over Fabric device with minimal hardware outside the flash storage hardware ...learn more

Networking

Overview / Usage

NVMe over fabric nodes are limited by the PCI bandwidth available to the CPU to move both network fabric traffic and NVMe data movement traffic to the storage subsystem.

With modern 100Gb fabrics we will need significant PCI bandwidth to scale up the end node, and good PCI switching to scale out the NVMe device traffic.

A solution will be to provide significant hardware bandwidth with minimal supporting operating system, memory overhead to allow devices to scale out as inexpensively as possible.

An end node will contain a pair of 100GB links to the fabric using 16 PCIeGen3 links, 5 PCIeGen3 x4 links to a PCIGen3 fabric allowing us to connect 40 TB of flash to each node.

The operating environment will need to be minimized to support the fabric, DPDK, and the Storage stack

Comments (0)