Re: [patch] sched: add locking when update the task_group's cfs_rq[] array.

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

On Wed, 2008-11-19 at 09:21 -0800, Ken Chen wrote:
> On Wed, Nov 19, 2008 at 8:54 AM, Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:
> > On Tue, 2008-11-18 at 22:48 -0800, Ken Chen wrote:
> >> add locking when update the task_group's cfs_rq[] array.  tg_shares_up()
> >> can be potentially executed concurrently on multiple CPUs with overlaping
> >> cpu mask depending on where task_cpu() was when a task got woken up.  Lack
> >> of any locking while redistribute tg->shares over cfs_rq[] array opens up
> >> a large window for conflict updates and utimately cause corruptions to the
> >> integrity of per cpu cfs_rq shares. Add a tg_lock to protect the operations.
> >
> > I see why you want to do this, but introducing a global lock makes me
> > sad :/
> 
> I wholly agree on the scalability.  The bigger the system, the more it
> needs to protect the integrity of cfs_rq[]->shares that the sum still
> adds up to tg->shares.  Otherwise, the share distributed on each CPU's
> cfs_rq might go wildly and indirectly leads to fluctuation of
> effective total tg->shares.  However, I have the same doubt that this
> will scale on large CPU system.  Does CFS really have to iterate the
> whole task_group tree?

Yes, sadly. The weight of a per-cpu super-task representation depends on
the group's task distribution over all tasks :/

(Dhaval, could you send Ken a copy of the paper we did on this?)

The idea was that we balance the stuff usng the sched-domain tree and
update it incrementally, and on the top level sched domain fix it all
up.

Will it scale, half-way, I'd say. It races a little, but should
converge. The biggest issue is that we're running with 10 bit fixed
point math, and on large cpu machines you get into granularity problems.