Abstract
NVMe® over Fabrics (NVMe-oF™) performance is compared using RDMA (iWARP & RoCEv2) and TCP transports across 100Gb Ethernet. Synthetic and real-world workloads are applied across different Fabrics to a storage target using both standard (1500B) and jumbo (9000B) MTU frames. Performance is compared between a six-SSD 3D XPointTM LUN and a 6-SSD 3D NAND LUN. Results show the impact of different workloads (synthetic corner case and real-world workloads), RDMA & TCP transport mechanisms (CPU onload vs CPU offload) and different types of storage LUNs (3D XPoint vs 3D NAND).