Re: [PATCH] sched: smpnice work around for active_load_balance()

[Siddha, Suresh B]

On Thu, Mar 30, 2006 at 10:40:24AM +1100, Peter Williams wrote:
> Siddha, Suresh B wrote:
> > On Wed, Mar 29, 2006 at 02:42:45PM +1100, Peter Williams wrote:
> >> I meant that it doesn't explicitly address your problem.  What it does 
> >> is ASSUME that failure of load balancing to move tasks is because there 
> >> was exactly one task on the source run queue and that this makes it a 
> >> suitable candidate to have that single task moved elsewhere in the blind 
> >> hope that it may fix an HT/MC imbalance that may or may not exist.  In 
> >> my mind this is very close to random.  
> > 
> > That so called assumption happens only when load balancing has
> > failed for more than the domain specific cache_nice_tries. Only reason
> > why it can fail so many times is because of all pinned tasks or only a single
> > task is running on that particular CPU. load balancing code takes care of both
> > these scenarios..
> > 
> > sched groups cpu_power controls the mechanism of implementing HT/MC
> > optimizations in addition to active balance code... There is no randomness
> > in this.
> 
> The above explanation just increases my belief in the randomness of this 
> solution.  This code is mostly done without locks and is therefore very 
> racy and any assumptions made based on the number of times load 
> balancing has failed etc. are highly speculative.

Isn't it the same case with regular cpu load calculations during load
balance?

> And even if there is only one task on the CPU there's no guarantee that
> that CPU is in a package that meets the other requirements to make the 
> move desirable.  So there's a good probability that you'll be moving 
> tasks unnecessarily.

sched groups cpu_power and domain topology information cleanly
encapsulates the imbalance identification and source/destination groups
to fix the imbalance.

> It's a poor solution and it's being inflicted on architectures that 
> don't need it.  Even if cache_nice_tries is used to suppress this 
> behaviour on architectures that don't need it they have to carry the 
> code in their kernel.

We can clearly throw CONFIG_SCHED_MC/SMT around that code.. Nick/Ingo
do you see any issue?

> 
> > 
> > 
> >> Also back to front and inefficient.
> > 
> > HT/MC imbalance is detected in a normal way.. A lightly loaded group
> > finds an imbalance and tries to pull some load from a busy group (which
> > is inline with normal load balance)... pull fails because the only task
> > on that cpu is busy running and needs to go off the cpu (which is triggered
> > by active load balance)... Scheduler load balance is generally done by a 
> > pull mechansim and here (HT/MC) it is still a pull mechanism(triggering a 
> > final push only because of the single running task) 
> > 
> > If you have any better generic and simple method, please let us know.
> 
> I gave an example in a previous e-mail.  Basically, at the end of 
> scheduler_tick() if rebalance_tick() doesn't move any tasks (it would be 
> foolish to contemplate moving tasks of the queue just after you've moved 
> some there) and the run queue has exactly one running task and it's time 
> for a HT/MC rebalance check on the package that this run queue belongs 
> to then check that package to to see if it meets the rest of criteria 
> for needing to lose some tasks.  If it does look for a package that is a 
> suitable recipient for the moved task and if you find one then mark this 
> run queue as needing active load balancing and arrange for its migration 
> thread to be started.
> 
> Simple, direct and amenable to being only built on architectures that 
> need the functionality.

First of all we will be doing unnecessary checks to see if there is
an imbalance.. Current code triggers the checks and movement only when
it is necessary.. And second, finding the correct destination cpu in the 
presence of SMT and MC is really complicated.. Look at different examples
in the OLS paper.. Domain topology provides all this info with no added
complexity...

> Another (more complex) solution that would also allow improvements to 
> other HT related code (e.g. the sleeping dependent code) would be to 
> modify the load balancing code so that all CPUs in a package share a run 
> queue and load balancing is then done between packages.  As long as the 
> number of CPUs in a package is small this shouldn't have scalability 
> issues.  The big disadvantage of this approach is its complexity which 
> is probably too great to contemplate doing it in 2.6.X kernels.

Presence of SMT and MC, implementation of power-savings scheduler
policy will present more challenges...

thanks,
suresh
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/