Rethinking Storage Software for Fast Storage Hardware

Recent advances in storage technologies have introduced devices with performance characteristics far beyond those of conventional storage systems. These developments challenge long-standing assumptions in the design of storage software and call for a fundamental rethinking of the storage stack.

This seminar presents two examples of research on Optane-based solid-state devices that illustrate both the need for and the benefits of redesigning storage systems for this new hardware landscape. The first example is KVell, a key-value store designed around the observation that on fast storage devices, system performance is often limited by CPU overhead rather than device latency. KVell therefore focuses on minimizing CPU intervention in the data path.

The second example is KVell+, an OLTP/OLAP system built on top of KVell. With the dramatically increased speed of modern storage devices, traditional approaches to snapshot isolation; relying on maintaining multiple data versions; can lead to excessive space amplification. KVell+ addresses this challenge by processing versions at creation time rather than storing them for later reconciliation.

This work is conducted in collaboration with researchers from academia and industry, including McGill University, Coupang, Nutanix, Google, INRIA, and the University of Sydney.