Re: [PATCH] posix-timers: RCU conversion

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Tue, 2011-03-22 at 12:35 +0200, Avi Kivity wrote:
>> Here's top with 96 idle guests running:

On some hacked up 2.6.38 kernel...

> > Start of perf report -g
> >      55.26%            kvm  [kernel.kallsyms]     [k] __ticket_spin_lock
> >                        |
> >                        --- __ticket_spin_lock
> >                           |
> >                           |--94.68%-- _raw_spin_lock
> >                           |          |
> >                           |          |--97.55%-- double_rq_lock
> >                           |          |          load_balance
> >                           |          |          idle_balance
> >                           |          |          schedule
> >                           |          |          |
> >                           |          |          |--60.56%-- schedule_hrtimeout_range_clock
> >                           |          |          |          schedule_hrtimeout_range
> >                           |          |          |          poll_schedule_timeout
> >                           |          |          |          do_select
> >                           |          |          |          core_sys_select
> >                           |          |          |          sys_select
> >                           |          |          |          system_call_fastpath

Looks like your workload and idle balancing don't much like each-other.

What I think is happening is that all your 'idle' qemu thingies keep
waking up frequently and because you've got like twice the number of
qemu instances as you've got cpus there's a fair chance you'll have a
cpu with a pending task while another one goes idle.

(Why does qemu keep waking if its idle? broken NOHZ?)

So idle balancing is called when the cpu goes idle (context switch to
the idle thread) and tries to steal a pending task from another cpu,
clearly it keeps finding these tasks otherwise it wouldn't try to take
that lock.

Mike, you build in some idle balance throttle logic, but that seems
defeated here (possible because it keeps finding pending tasks to
migrate? still need morning juice).