Bridging CPU and GPU I/O with a Device-Resident File System

Abstract

Modern AI and data analytics workloads have terabyte-scale datasets, making GPUs access storage frequently. GPUDirect Storage allows the data path from SSD to GPU to bypass CPU memory. However, the file system shared by the CPU and GPU still runs on the CPU. Existing GPU file systems cache metadata to avoid frequent accesses to the CPU. However, any update to file metadata, such as file length or names, must be synchronized between both sides. We argue that frequent metadata synchronization becomes a bottleneck in such systems. A device-resident file system can bridge CPU-initiated and GPU-initiated I/O by maintaining a single authoritative metadata instance, thereby eliminating frequent CPU–GPU metadata synchronization. This paper presents DevMeta, a device-resident file system that realizes this idea, eliminates the need for frequent CPU–GPU metadata synchronization, and explores the key challenges of implementing a file system at the device level.