SingularFS: A Billion-Scale Distributed File System Using a Single Metadata Server

Billion-scale distributed file systems play an important role in modern datacenters, and it is desirable and possible to support these file systems with a single metadata server. However, fully exploiting its performance faces unique challenges, including crash consistency overhead, lock contention in a shared directory, and NUMA locality.

This paper presents SingularFS, a billion-scale distributed file system using a single metadata server. It includes three key techniques. First, SingularFS proposes log-free metadata operations to eliminate additional crash consistency overheads for most metadata operations. Second, SingularFS designs hierarchical concurrency control to maximize the parallelism of metadata operations. Third, SingularFS introduces hybrid inode partition to reduce inter-NUMA access and intra-NUMA lock contention. Our extensive evaluation shows that SingularFS consistently provides high performance for metadata operations on both private and shared directories, and has a steadily high throughput for the billion-scale directory tree.