Re: [RFC][PATCH 0/5] Reduce runqueue lock contention -v2

[Frank Rowand <frank.rowand@gmail.com>]

On 12/16/10 06:56, Peter Zijlstra wrote:
> Hi, here a new posting of my scary patch(es) ;-)
> 
> These actually survive a sembench run (and everything else I threw at it).
> The discussion between Mike and Frank over the task_running() check made me
> realize what was wrong with the previous one.
> 
> As it turns out, what was needed (p->oncpu) was something Thomas wanted me
> to do for an entirely different reason (see patch #2).
> 
> Frank's patch, while encouraging me to poke at it again, has a number of
> very fundamental problems with it, the most serious one being that it
> completely wrecks the wake-up load-balancing.

And also as Peter pointed out when I posted the patch (thank you Peter),
I did not properly handle the return value for try_to_wake_up() - a rather
fatal flaw.

By coincidence, I was about to post a new version of my scary patch when
this email arrived.  I'll post my patches as a reply to this email, then
read through Peter's.


Frank's patch, Version 2

Changes from Version 1:
  - Ensure return value of try_to_wake_up() is correct, even when queueing
    wake up on a different cpu.
  - rq->lock contention reduction not as good as first version

patch 1

  The core changes.  All the scary lock related stuff.

  select_task_rq() uses the smp_processor_id() of the old task_cpu(p) instead
  of the waking smp_processor_id().

patch 2

  select_task_rq() uses the smp_processor_id() of the waking smp_processor_id()
  
Limitations
  x86 only

Tests
  - tested on 2 cpu x86_64
  - very simplistic workload
  - results:
     rq->lock contention count reduced by ~ 75%
     rq->lock contention wait time reduced by ~ 70%
     test duration reduction is in the noise
     rq->lock contention improvement is slightly better with just patch 1
       applied, but the difference is in the noise

Todo
  - handle cpu being offlined


-Frank