Over the years, the storage substrate of operating systems has evolved with changing storage devices and workloads [2, 6, 7, 8, 12, 15, 18, 26, 29, 33, 34, 35, 39, 41, 42, 44, 47, 48, 54]. Both academia and industry have devoted significant research effort to the file system component, a critical part of the storage system. A file system directs the underlying device-specific software to perform data reads and writes as well as providing the notion of files to interact with users and applications. To achieve this, a file system represents logical files internally or physically with data (the file content) and metadata (information required to locate, index, and operate on data). Most file system optimizations assume this one-to-one coupling of logical and physical representations [2, 7, 8, 18, 25, 26, 29, 33, 34, 35, 48]. This dissertation presents the design, implementation, and evaluation of two new systems, which decouple these representations and offer a new class of optimization opportunities not previously possible. First, the Composite-File File System (CFFS) exploits the observation that many files are frequently accessed together. By consolidating related file metadata, performance can be improved by up to 27%. Second, the Fine-grained Journal Store (FJS) exploits the observation that typically only subregions of a metadata entry are updated, but the heavyweight reliability and storage mechanisms then affect the entire metadata entry. This results in many unnecessary metadata writes that harm both the performance and the lifespan of certain storage devices. By focusing on only the updated metadata regions and consolidating storage and reliability mechanisms, the Fine-grained Journal Store can both improve the performance up to 15x and reduce unnecessary writes up to 5.8x. Overall, the decoupling of logical and physical representations allows more flexible matching of the physical representations to the workload patterns, and the results show that this approach is promising.
