Uploaded image for project: 'Lustre'
  1. Lustre
  2. LU-20 patchless server kernel
  3. LU-2498

set block device IO scheduler to "deadline"

    XMLWordPrintable

Details

    • Technical task
    • Resolution: Fixed
    • Minor
    • Lustre 2.5.0
    • Lustre 2.4.0
    • 5855

    Description

      If we are moving to a patchless server kernel, Lustre will need to automatically set the block device scheduler to "deadline" if it is currently set to "cfq", since the "cfq" scheduler exhibits very poor performance for high-bandwidth applications like Lustre. We shouldn't override the setting if "noop" is set by the administrator.

      Currently this is done via the kernel config file setting CONFIG_DEFAULT_DEADLINE=y in our config files, but this will not be possible if we are not building our own kernels (the upstream default is "cfq", which is more suitable for desktop use).

      This should probably be done via mount_utils_ldiskfs.c for ldiskfs devices (since the device name is specified on the command-line). This is more difficult for ZFS devices, since there is not a simple mapping from the mounted "device" and the underlying disks. I suspect it should be done internal to the "zfs" utility (or in the kernel) itself, when the device is first being imported.

      Attachments

        Activity

          People

            prakash Prakash Surya (Inactive)
            adilger Andreas Dilger
            Votes:
            0 Vote for this issue
            Watchers:
            6 Start watching this issue

            Dates

              Created:
              Updated:
              Resolved: