Paper Title: Accelerating NVMe over Fabrics with Hardware Offloads at 100Gb/s and Beyond

Paper Abstract: A new wave of NVMe over Fabrics solutions are reaching market, but as SSDs and networks get faster, CPUs are having trouble keeping up. Specifically the traditional data path of routing NVMe traffic through the CPU and main memory begins to present throughput and latency bottlenecks on the storage target side. The solution is a new class of intelligent network adapters and SOCs (systems on chip) that can offload not only the RDMA (Remote Direct Memory Access) and networking, but also offload the NVMe-oF protocol itself. By processing the NVMe commands within the network adapter or SOC, extra interrupts to the CPU can be avoided. By writing data directly to the SSDs or using local cache within the SOC, the main memory can be bypassed. Both techniques reduce latency and increase the number of IOPS, which becomes more critical as network speeds stretch from 100Gb/s up to 200Gb/s.

Paper Author: Idan Burstein, Staff Silicon Architect, Mellanox

Author Bio: Oren Duer is Principal Software Architect at Mellanox Technologies, where he focuses on developing storage accelerations and offloads. He works to enable those features in different operating systems, hypervisors, and storage software stacks. Oren is currently leading the development of NVMe over Fabrics and Persistent Memory over Fabrics accelerations, and previously led the development of RDMA (Remote Direct Memory Acceleration) in other storage protocols and solutions, including iSER (iSCSI Extensions for RDMA), SRP (SCSI RDMA Protocol), and Ceph. Previously Oren was Director of Storage Software for Mellanox and prior to that worked as a software engineer at Montilio. Oren has an computer science degree from The Hebrew University of Jerusalem.