Prof. Dr. Frank Bellosa

Yussuf Khalil

Abstract
The emergence of CXL-based persistent storage, which can be accessed both asynchronously through CXL.io and synchronously through CXL.mem, has led to the development of Transparent DAX Mappings (TDMs). TDMs allow applications to take advantage of CXL’s Direct Access (DAX) feature without requiring modifications. They work by transparently modifying read and write calls to use DAX mappings for I/O. However, supporting traditional storage features such as RAID in this context is challenging due to the desire to leave the kernel out of the data path.In this thesis, we present CXLRaid, a userspace-only implementation of RAID 1 and RAID 5 that extends the TDM concept to provide RAID functionality on a file level. It leverages files on multiple devices as storage backends, each of which is accessed through a DAX mapping. This provides the necessary redundancy for RAID while not involving the kernel in the data path at all. With this approach, we achieve up to 3.8× higher write bandwidth for RAID 5 and up to 1.4× higher read bandwidth for RAID 1 compared to the Linux kernel’s md/raid implementation, while maintaining comparable performance for RAID 1 writes. We also discuss possible future work on our implementation, as well as propose a RAID design that continues to use TDMs for data transfer but moves other logic into the kernel for closer integration.

BibTex:

@bachelorthesis{dierheimer26SoftwareRAID,
  author = {Linus Paul Dierheimer},
  title = {Software RAID for CXL-Based SSDs },
  type = {Bachelor Thesis},
  year = 2026,
  month = mar# "26",
  school = {Operating Systems Group, Karlsruhe Institute of Technology (KIT), Germany}
  }