Paper Title: FPGA Accelerator Disaggregation using NVMe-over-Fabrics

Paper Abstract: NVMe-over-Fabrics (NVMe-oF) is a network protocol for NVMe that is used to communicate between a host machine and an NVMe based storage system over high-performance Ethernet networks. Standard NVMe devices can leverage NVMe-oF to share namespaces across existing network fabrics such as RDMA, TCP/IP, or Fibre Channel. This talk will discuss a method to leverage the NVMe standard to present FPGA accelerators as NVMe namespaces to the operating system, thus allowing FPGA accelerators to be accessed with simple block device read/write operations via in-box drivers. With NVMe-oF, client machines can simply borrow namespaces/accelerators over the fabrics connection which it can then see and access as if it were local to the system. Because server nodes can share these accelerators on demand, this setup allows for the disaggregation of acceleration compute resources. With the addition of a Controller Memory Buffer (CMB), the fabrics connection device can send/receive data directly to/from a FPGA accelerator without accessing any host memory. This talk will also discuss the advantages of this disaggregation, as well as its impact on resource usage.

Paper Author: Andrew Maier, Software Engineer, Eideticom

Author Bio: Andrew Maier has been involved in the NoLoad™ project as a Software Engineer with Eideticom since January 2017. He has a B.Sc. and M.Sc. in Computer Engineering from the University of Alberta and his recent research included the acceleration of LDPC codes using OpenCL for FPGAs. He participated in the international robotics competition BattleBots as part of team MBSRobotics that aired on ABC in the summer of 2016.