Improving file system consistency and durability with patches and bpfs

This dissertation improves the consistency and durability guarantees that file systems can efficiently provide, both by allowing each application to choose appropriate trade-offs between consistency and performance and by dramatically lowering the overheads of durability and consistency using new hardware and careful file system design. We first describe a new abstraction, the patch, which represents a write to persistent storage and its ordering requirements. This abstraction allows file system modules to specify ordering guarantees without simultaneously requesting more expensive, immediate durability. Algorithmic and data structure optimizations make this abstraction practical, and our patch-based file system implementation is performance-competitive with similarly-reliable Linux ext2 and ext3 configurations. To illustrate the benefits of patchgroups, an application-accessible version of patches, we apply them to improve the performance of the UW IMAP server by over 13 times and to make file handling in gzip robust to crashes. In the second part of this work we investigate using upcoming byte-addressable, persistent memory technologies—in particular, phase change memory—in place of disks and flash to reduce the costs of enforcing ordering constraints and providing durability and atomicity. We describe a new file system, BPFS, that commits each file system operation synchronously and atomically. BPFS exploits byte-addressability, improved throughput and latency, and our new atomic write primitive to eliminate copy-on-writes that have until now been required to implement shadow paging. Our evaluation shows that, because of our optimizations, BPFS provides its exceptionally stronger guarantees on phase change memory without lowering the throughput that today's file systems achieve on disks.