cvs commit: src/sys/kern sched_4bsd.c

[Bruce Evans]

bde         2003/11/09 05:45:55 PST

  FreeBSD src repository

  Modified files:
    sys/kern             sched_4bsd.c 
  Log:
  Quick fix for scaling of statclock ticks in the SMP case.  As explained
  in the log message for kern_sched.c 1.83 (which should have been
  repo-copied to preserve history for this file), the (4BSD) scheduler
  algorithm only works right if stathz is nearly 128 Hz.  The old
  commit lock said 64 Hz; the scheduler actually wants nearly 16 Hz
  but there was a scale factor of 4 to give the requirement of 64 Hz,
  and rev.1.83 changed the scale factor so that the requirement became
  128 Hz.  The change of the scale factor was incomplete in the SMP
  case.  Then scheduling ticks are provided by smp_ncpu CPUs, and the
  scheduler cannot tell the difference between this and 1 CPU providing
  scheduling ticks smp_ncpu times faster, so we need another scale
  factor of smp_ncp or an algorithm change.
  
  This quick fix uses the scale factor without even trying to optimize
  the runtime divisions required for this as is done for the other
  scale factor.
  
  The main algorithmic problem is the clamp on the scheduling tick counts.
  This was 295; it is now approximately 295 * smp_ncpu.  When the limit
  is reached, threads get free timeslices and scheduling becomes very
  unfair to the threads that don't hit the limit.  The limit can be
  reached and maintained in the worst case if the load average is larger
  than (limit / effective_stathz - 1) / 2 = 0.65 now (was just 0.08 with
  2 CPUs before this change), so there are algorithmic problems even for
  a load average of 1.  Fortunately, the worst case isn't common enough
  for the problem to be very noticeable (it is mainly for niced CPU hogs
  competing with less nice CPU hogs).
  
  Revision  Changes    Path
  1.28      +4 -0      src/sys/kern/sched_4bsd.c