Prof. Dr. Frank Bellosa
Peter Maucher

This thesis presents a flexible RAID system integrated into GPU4FS, a novel GPU accelerated file system. Software RAID systems provide flexibility over hardware RAID systems, yet tasks involving complex parity coding strain CPUs, especially in double and triple parity configurations. To address this, we propose a GPU based RAID system integrated into GPU4FS, leveraging the GPUs parallel processing for efficient parity calculation and data handling.
Our core design concept centers on a logical address space managed by GPU4FS, allowing dynamic allocation of files and pages with specific RAID configurations, tailoring performance and redundancy to individual files and processes. Evaluations across RAID levels consistently demonstrated higher write bandwidths and reduced CPU utilization compared to CPU based RAID systems on Optane. Furthermore, the evaluation showed that parity calculation adds minimal computational overhead, affirming the efficiency of our GPU centric approach.
By harnessing GPU computational power, we present an innovative alternative to CPU based RAID systems, offering competitive bandwidths while reducing the average CPU usage by a factor of x13 to x21 depending on the RAID level.

BibTex:

@bachelorthesis{lucka23RAIDGPU4FS,
  author = {Gregor Lucka},
  title = {RAID on a File System Level for GPU4FS},
  type = {Bachelor Thesis},
  year = 2023,
  month = sept # "18",
  school = {Operating Systems Group, Karlsruhe Institute of Technology (KIT), Germany}
  }