Re: [LSF/MM TOPIC] Native NVMM file systems

PMEM/DAX should allow for significant improvements in file system performance and enable new programming models that allow direct, efficient access to PMEM from userspace. Achieving these gains in existing file systems built for block devices (e.g., XFS and EXT4…) presents a range of challenges (e.g., https://lkml.org/lkml/2016/9/11/159) and has been the subject of a lot of recent work on ext4 and xfs. An alternative is to build a NVMM-aware file system from scratch that takes full advantage of the performance that PMEM offers and avoids the complexity that block-based file systems include to maximize performance on slow storage (e.g., relaxing atomicity constraints on many operations). Of course, it also brings with it the complexity of another file system. We recently sent out a patch set for one-such “clean slate” NVMM-aware file system called NOVA. NOVA is log-structured DAX file system with several nice features: * High performance, especially in metadata operations due to efficient fine-grained logging * High scalability with per-CPU memory pool and per-inode logging * Strong metadata and data atomicity guarantees for all operations * Full filesystem snapshot support with DAX-mmap * Metadata replication/checksums and RAID-4 style data protection At the summit, we would like to discuss the trade-offs between adapting NVMM features to existing file systems vs. creating/adopting a purpose-built file system for NVMM. NOVA serves as useful starting point for that discussion by demonstrating what’s possible. It may also suggest some features that could be adapted to other file systems to improve NVMM performance. We welcome people that are interested in file systems and NVM/DAX. Particular people that would be useful to have in attendance are Dan Williams, Dave Chinner, and Matthew Wilcox.