[prev in list] [next in list] [prev in thread] [next in thread] 

List:       linux-ia64
Subject:    [Linux-ia64] O(1) scheduler for 2.4.18-ia64
From:       Erich Focht <efocht () ess ! nec ! de>
Date:       2002-07-16 16:23:52
[Download RAW message or body]

Hi,

for those working with IA64 systems: at
  http://home.arcor.de/efocht/sched/MQO1_ia64-ef7-2.4.18.patch.bz2
you can find a port of the O(1) scheduler. It is the "old" one, i.e.
not having SCHED_BATCH, etc. But it has some important changes compared
to my last port:

  - Uses the "complex" macros which release the RQ lock before switching
    context. This is a MUST on IA64, as there is a potential deadlock
    when the context numbers are wrapped around.

  - cpu_idle() is changed such that platforms which just return from
    pm_idle() don't call schedule() in a tight loop. This was the reason
    for massive cache line bouncings when some CPUs were idle as they
    not only locked their RQs but also loaded nr_running (on the same cac=
he
    line as the lock) for every other RQ. The position of nr_running insi=
de
    the RQ structure is changed, such that it doesn't share the cacheline
    with lock.

  - load_balance() only steals at most one task. Before it could sometime=
s
    steal half of the tasks of the most loaded CPU. This is OK on small
    machines but on large ones the tasks were bounced around a while unti=
l
    "settling" on a particular CPU. Maybe this is not yet optimal, but it
    improves hackbench results.

The patch applies to  2.4.18  +  4th ia64 patch from David. It also can
be applied to the vanilla 2.4.18, (kernel/timer.c needs to be edited by
hand), but this port is not intended to be an alternative to Robert Love'=
s
backports... It is for IA64 and/or the node affine scheduler.

Best regards,

Erich



[prev in list] [next in list] [prev in thread] [next in thread]