Oxbow: A Coordinated Architecture for Multi-Component File Systems

Fast storage hardware and computational SSDs have outpaced the traditional kernel-centric or kernel-bypass file system designs, fragmenting modern storage stacks across library file systems, kernel subsystems, and in-device file systems. Each architecture offers only a subset of desired properties: userlevel designs deliver low latency but lose kernel services and isolation, kernel file systems retain rich functionality but become CPU-bound and slow, and device-resident logic reduces host load but suffers from PCIe latency and wimpy processors.

This paper presents Oxbow, a coordinated storage architecture that composes kernel, user-space, and device components to achieve all four goals simultaneously: high performance, strong kernel interoperability, low CPU consumption, and fast development velocity. Oxbow combines a kernel-based read path with a kernel-bypassing write path, shared-ownership metadata, and Split Journaling, a host–device journaling mechanism that decouples fsync from background commits using staging areas and DMA-based snapshots. We demonstrate that Oxbow delivers kernel-level protection and sharing semantics while matching or exceeding the performance of state-of-the-art user-level file systems and substantially reducing host CPU use through offload to computational SSDs.