Extending GPU4FS with Advanced File System Functionalities

  • Type:Bachelor Thesis
  • Date:29.08.2023
  • Supervisor:

    Prof. Dr. Frank Bellosa
    Peter Maucher

  • Graduand:Nico Rath
  • Links:PDF
  • Persistent non-volatile memory (PMem) gained more and more attention over the years. When compared to SSDs, it shines with lower latency and higher read throughput. However, its write bandwidth is a major drawback. As PMem is connected via the memory bus, the OS does not detect an IO-bound task; which lets CPU cores stall more likely. Shifting PMem interaction to an accelerator is one way to resolve this problem. GPU4FS is a graphics card accelerated file system. However, it does not provide any advanced features. In this thesis, we extend GPU4FS with checksum and deduplication functionality. While checksums protect a file system’s content against corruption, deduplication can save storage by detecting duplicate data blocks. Both functionalities are suitable for a GPU, as their implementation is parallelizable. Our results show that GPU acceleration of
    those features becomes more feasible the more load the file system faces. Thus, a hybrid mode with mixed CPU/GPU utilization seems like a promising concept.

    BibTex:

    @bachelorthesis{rath23extendingGPU4FS,
      author = {Nico Rath},
      title = {Extending GPU4FS with Advanced File System Functionalities},
      type = {Bachelor Thesis},
      year = 2023,
      month = aug # "29",
      school = {Operating Systems Group, Karlsruhe Institute of Technology (KIT), Germany}
      }