BIFOLD Colloquium 09/2025

How to Build an SSD

Abstract: Database system designers aim at leveraging NVMe SSDs for overall performance and sustainability. I will start this talk by reviewing recent work, led by Gabriel Haas, showing that the choice of SSDs matter in this respect. Indeed, SSD models that have the same capacity and the same underlying NAND flash technology (e.g., MLC or TLC) embody varying undocumented trade-offs between cost, energy consumption and performance. Today, database designers and platform engineers must choose appropriate SSDs with care. But this is far from ideal. We posit that, tomorrow, database designers should be able to build SSDs that match their needs and provide appropriate end-to-end trade-offs. In the CHORYS project, we are designing OpenSSD-V for this purpose. In this talk, I will detail the motivation for this work, our approach as well as early results. OpenSSD-V is joint work with Alberto Lerner.

Bio: Philippe Bonnet is professor in the Department of Computer Science at the University of Copenhagen (DIKU). He is an experimental computer scientist with a background in database systems. He is an expert in storage systems software. In this field, he contributed to the uFlip Benchmark, the Linux multiqueue block layer, the Linux framework for Open-Channel SSDs, the OX architecture for computational storage, the xNVMe library, and Delilah, a prototype for eBPF offload on computational storage. Philippe was involved in the development of sensor networks in the early 2000s with the Cougar project at Cornell and the Hogthrob and Mana projects at the University of Copenhagen. Philippe chairs the ACM EIG on Reproducibility and Replicability. He is the co-author of a book on the Principles of Database and SSD Co-Design. He coordinates the CHORYS project, funded by the European Union.