Managing Very Large Data Sets on Directly-Attached NVMe Arrays

Managing Very Large Data Sets on Directly-Attached NVMe Arrays (LE3948/1)

PI: Viktor Leis (Friedrich-Schiller-Universität Jena)
Project Collaborator: Gabriel Haas
Project website: http://leanstore.io

PCIe-attached solid-state drives offer high throughput and large capacity at low cost. Modern servers can easily host 4, 8, or 16 such SSDs, resulting in an aggregated bandwidth that hitherto has only been achievable using DRAM. In this project, we investigate how to exploit Directly-Attached NVMe Arrays (DANA) in high-performance storage engines. This requires revisiting how I/O operations are scheduled by the storage engine, how the operating system handles hundreds of concurrent I/O requests, and how to take the physical characteristics of NAND flash into account.