No announcement yet.

The "Memory Folios" Work Continues - Improving Linux Performance, 7% Faster Kernel Builds

  • Filter
  • Time
  • Show
Clear All
new posts

  • #11
    A few years back, a NAS I worked on was using an ARM chip that could work with either 4K or 64K pages. Between the slowness of the CPU (~600MHz), and the overhead of the massive page tables needed to hold GbE-worth of data in-flight to/from an HDD, the change to the 16x larger pages netted something like a 15% increase in performance, finally allowing the woefully-underresourced thing (it had basically half the RAM it "should have" had) to be able to actually sustain line rate rather than constantly drooping (sic) down to 800-ish Mb/s thanks to all the extra work the 4K build had to do.

    So while this really only "matters" for the ends of the spectrum - i.e. weak ARM devices and powerful servers - rather than any typical desktop case, it genuinely can matter for at least those extremes, and this patchset is very much a Good Thing in its own right, even without counting the code cleanup etc.
    Last edited by arQon; 17 June 2021, 09:15 PM.


    • #12
      Originally posted by nsneck View Post
      Interesting and impressive work. Of course looking forward to many more patchsets, with this maybe landing next year.

      Why does this require separate patches to XFS, and why has this only been done for XFS? From the patches this seems to touch just the kernels memory management side, do filesystems really need separate logic for new MM paths?
      Both the author of the memory folio work, and the core maintainer for XFS, work for Oracle. XFS is a nice and mature kernel component, probably makes for a logical proof of concept.