git: kernel - Fix several usched nits

[Matthew Dillon]

commit cbd77b01f540acf8bda7ccb769365f5d4cb577e4
Author: Matthew Dillon <dillon at apollo.backplane.com>
Date:   Thu Apr 26 23:11:07 2018 -0700

    kernel - Fix several usched nits
    
    * Fix an issue where a usched-restricted process is pulled onto a
      cpu that is outside of its mask.  The process will immediately
      push itself back onto a cpu in its mask, but this bug leads to
      unwanted cpu ping ponging.
    
    * Fix an issue where cpu-bound usched-restricted processes can
      cause other processes to be misscheduled at a lower priority
      than their actual priority, resulting in unexpected stuttering.
    
    * With these changes, nice +/- priorities should work considerably
      better.  In the normal case, alotted cpu time will be as shown
      below.  This is intended to be non-linear:
    
         1   RQ0N    1:42.11 forever
         2   R0N     1:22.55 forever
         3   R0N     0:51.46 forever
         4   RQ0N    0:24.23 forever
         5   R0N     0:23.12 forever
         6   RQ0N    0:19.91 forever
         7   RQ0N    0:09.08 forever
         8   RQ0N    0:08.38 forever
         9   RQ0N    0:06.46 forever
        10   RQ0N    0:06.72 forever
        11   RQ0N    0:04.65 forever
        12   RQ0N    0:04.60 forever
        13   RQ0N    0:03.04 forever
        14   RQ0N    0:02.53 forever
        15   RQ0N    0:02.16 forever
        16   RQ0N    0:01.55 forever
        17   RQ0N    0:01.25 forever
        18   RQ0N    0:00.81 forever
        19   RQ0N    0:00.76 forever
        20   RQ0N    0:00.32 forever
    
    * In addition, dynamic priority operation should also work better with
      these changes.
    
    * WARNING!  With this change, nice values actually work as intended.  A
      nice +20 process will get very little cpu in the face of a cpu-bound
      nice +0 process on a cpu.  Similarly, a cpu-bound nice -20 process will
      allow nice +0 processes to barely function on a cpu.
    
      And if you use an even larger spread, a cpu-bound nice -20 process
      will not allow a nice +1 or higher process any cpu time at all.
    
      Having multiple cores mitigations this somewhat, but users must be
      careful especially when specifying negative nice values for processes.
    
    * If you run X and have jerkiness issues due to process load, you may want
      to give the X server itself around a nice -5.  You can fiddle, but in
      basic testing with everything running at nice 0, the scheduler appears
      to do a pretty good just generally.  For example, if your browser is
      so overloaded that it has essentially become cpu bound, it will get cpu
      on an equal basis with batch jobs like bulk builds.

Summary of changes:
 sys/kern/usched_dfly.c | 57 ++++++++++++++++++++++++++++++++++----------------
 1 file changed, 39 insertions(+), 18 deletions(-)

http://gitweb.dragonflybsd.org/dragonfly.git/commitdiff/cbd77b01f540acf8bda7ccb769365f5d4cb577e4


-- 
DragonFly BSD source repository