Paper Title: Selecting the Right Architecture for an All-Flash NVMe Solution

Paper Abstract: All-flash solutions can utilize a variety of architectures. At one end, a disaggregated solution can simply have a bunch of NVMe flash in a box with a PCIe switch, which is then connected to a group of servers. Smart software attaches individual flash devices to individual servers for as long as they are needed. Flash can thus be moved from one server to another as needed by applications, thus providing flexible assignment of high-performance dedicated storage. An individual box, for example, could consist of a 1U JBOF with 32 NVMe drives (utilizing the Ruler or EDSFF form factor). At the other end, a full networked solution can include servers and flash storage everywhere with access determined by the network operating system. Trade-offs include throughput, cost, scalability, number of PCIe lanes used, average and tail latency, and network speed. Obviously, the requirements of particular customer applications vary widely and determine which architecture is best for them.

Paper Author: Mike Scriber, Sr Director Server Solutions Management, SuperMicro

Author Bio: Mike Scriber is Sr Director Server Solution Management at Supermicro, where he designs very high density storage systems using NVMe. His recent work includes 1U racks with close to 1 petabyte capacity using U.2 drives and the Ruler (EDSFF) form factor. Before joining Supermicro, he was Director of Systems Engineering at Samsung Semiconductor, where he designed and developed the first 2U 40 NVMe SSD storage systems. He has also held senior engineering positions at Dell, Toshiba, IBM, and Intel during his 30-year career in the technology industry. He earned an MS in computer engineering from the Air Force Institute of Technology and a BS in computer engineering from Rensselaer Polytechnic Institute (RPI). He has also presented at SNIA’s Storage Developer Conference and Supermicro Build Day Live.