Paper Title: Optimizing Hyperconverged Infrastructure for NVMe-Based Flash Storage

Paper Abstract: The optimal choice for hyperconverged infrastructure has changed now that NVMe and NVMe-oF have become the most common flash storage connections. The traditional monolythic SAN-in-a-VM model is facing serious challenges as the bandwidth required for attached SSDs has increased greatly over what was typical for disk interfaces. The new generation of HCI software stacks, based on approaches such as replicated container-attached storage, has proven to be better suited for accommodating the unprecedented scale, density, and diversity of workloads. This however puts an increased strain on processor, interconnect, and network resources, which previously could readily handle large numbers of drives. Now infrastructure designers must consider new choices such as increasing the capabilities of central processors, using higher-bandwidth connections, or offloading functions to additional hardware such as FPGAs or coprocessors. All these options add to system cost and complexity. The correct choices depend on application areas, such as analytics, AI/ML, IoT, database searches, transactions processing, or video and image processing.

Paper Author: Kais Belgaied, Storage & Servers Division CTO, Sanmina

Author Bio: Kais Belgaied is currently CTO for Sanmina's Storage and Servers division, driving the company's adoption and growth strategy around NVMe. A prolific innovator with over 50 patents, he was previously the architect of the control plane scale-out and the availability of all-flash array clusters at the highly regarded startup Nimble Storage (acquired by HPE). He has also been an architect and team leader at VMware and Sun Microsystems. He is a self-starter with a proven record of successful project delivery in areas such as kernel development, software-defined storage, systems and networking virtualization, network protocols, and OS security and access controls. He has published articles and given or contributed to conference presentations on topics such as cache management, scalability, and hardware offload. He earned an MS and a postgraduate degree in computer science and applied mathematics from the Polytechnique Institute of Grenoble (France), and an MBA from Columbia.