Operating System Support for High-Performance Solid State Drives

The performance of Solid State Drives (SSD) has evolved from hundreds to millions of I/Os per second in the past three years. Such a radical evolution is transforming both the storage and the software industries. Indeed, software designed based on the assumption of slow IOs has become the bottleneck in the era of high-performance SSDs. In the context of the CLyDE project, a collaboration between the IT University of Copenhagen and INRIA Rocquencourt, we recognized the mismatch between software designed for hard drives and highperformance SSDs. This thesis focuses on the role of the operating system in reducing the gap, and enabling new forms of communication and even co-design between applications and high-performance SSDs. More specifically, we studied the storage layers within the Linux kernel. We explore the following issues: (i) what are the limitations of the legacy block interface and its cost in the context of high-performance SSDs? (ii) Can the Linux kernel block layer cope with high-performance SSDs? (iii) What is an alternative to the legacy block interface? How to explore the design space? (iv) How about exposing SSD characteristics to the host in order to enable a form of application-SSD co-design? What are the impacts on operating system design? (v) What would it take to provide quality of service for applications requiring millions of I/O per second? The dissertation consists of six publications covering these issues. Two of the main contributions of this thesis (the multi-queue block layer and LightNVM) are now an integral part of the Linux kernel.