Prof. Dr. Frank Bellosa
Lukas Werling

Abstract
Modern file systems with fast storage media take up valuable CPU resources, especially if used with Intel Optane in write-heavy applications. Optane is written synchronously, and accessing CPU cores stall if Optanes relatively low write bandwidth is saturated. To combat that issue, we propose GPU4FS, a novel GPUaccelerated user- space file system mainly targeted at Intel Optane as the main representative of high-performance non-volatile memory. GPU4FS relieves the CPU from all file system management tasks, including the writes to the Optane DIMMs. The CPU only queues requests to the file system into a shared command buffer.
With our prototype, we achieve similar bandwidth figures to CPU-side Optane tasks using a fio benchmark and sequential writes to DAX-EXT4. During the benchmark, the CPU tasks saturate four cores. Our GPU4FS prototype only uses a measured 12% of a single core, and offers a 33x reduction in CPU usage. We also show that we offer twice or more performance improvement for a parallel, CPU-bound blender render benchmark when running in parallel to GPU4FS, as compared to the CPU-side file system tasks.
Based on our findings, we conclude that future research into file system accelerators poses a worthwhile topic.

BibTex:

@masterthesis{maucher22gpu4fs,
  author = {Peter Maucher},
  title = {GPU4FS: A Graphics Processor-Accelerated File System},
  type = {Master Thesis},
  year = 2022,
  month = aug # "06",
  school = {Operating Systems Group, Karlsruhe Institute of Technology (KIT), Germany}
  }