Analyzing Shared Memory Opportunities in Different Workloads

  • Type:Study Thesis
  • Date:21.11.2011
  • Supervisor:

    Prof. Dr. Frank Bellosa, Konrad Miller

  • Graduand:Thorsten Gröninger
  • Links:PDF
  • Abstract:

    This thesis analyzes different workloads and their impact on memory sharing, sharing opportunities, and memory merging. To make the analysis possible, we have written a Linux kernel module to extract the needed information. It reduces memory dumps to 1.5% of the full memory dump size by hashing the page's content. These dumps show unused potential of page cache sharings between host and guest operating systems. Limitations of Kernel SamePage Merging as present in recent Linux kernels.
    We analyzed parallel kernel builds in virtual machines with activated KSM and discovered that up to 60% of sharing opportunities remain unmerged, due to the slow scanning rate of KSM. In this scenario up to 80% of data is redundant and could be shared, because 54% of these sharing opportunities last longer than 20 seconds, even in VMs with high memory pressure. Most of this sharing potential originates from the page cache, which also contains redudant pages (up to 10%). A more sophisticated deduplicating file system and page merging approach could speed up the whole system.


    author = {Thorsten Gr\"oninger},
    title = {Analyzing Shared Memory Opportunities in Different Workloads},
    type = {Study Thesis},
    school = {System Architecture Group, Karlsruhe Institute of Technology (KIT), Germany},
    month = nov # "21",
    year = 2011,
    note = {}