Achieving Optimal Throughput for Persistent Memory with Per-Process Accounting

  • Type:Master Thesis
  • Date:12.05.2022
  • Supervisor:

    Prof. Dr. Frank Bellosa
    Lukas Werling

  • Graduand:Thomas Schmidt
  • Links:PDF
  • Abstract
    In recent years non-volatile main memory (NVMM) emerged as a technology
    for byte-addressable persistent storage accessible similar to DRAM via the CPUs
    memory bus. Operating systems followed by introducing direct access (DAX)
    allowing applications to map NVMM into their address space, bypassing the OS
    on accesses.
    We contribute an NVMM usage monitor based on processor event-based sampling
    (PEBS) capable of detecting accesses to NVMM performed by applications.
    This can be used to interpolate an application’s NVMM utilization and perform
    system-wide per-process accounting.
    Further, prior work has shown Optane DC’s throughput to drop with the increasing
    number of parallel accesses. By limiting the set of schedulable CPU
    cores for threads that recently accessed NVMM, we also aim to improve the
    throughput at high numbers of concurrent threads.


    @masterthesis{schmidt22Per-Process Accounting,
      author = {Thomas Schmidt},
      title = {Achieving Optimal Throughput for Persistent Memory with Per-Process Accounting},
      type = {Master Thesis},
      year = 2022,
      month = may # "12",
      school = {Operating Systems Group, Karlsruhe Institute of Technology (KIT), Germany}