Re: [patch] sched: align rq to cacheline boundary

[Ingo Molnar <mingo@elte.hu>]

* Andrew Morton <akpm@linux-foundation.org> wrote:

> > -static DEFINE_PER_CPU(struct rq, runqueues);
> > +static DEFINE_PER_CPU(struct rq, runqueues) ____cacheline_aligned_in_smp;
> 
> Remember that this can consume up to (linesize-4 * NR_CPUS) bytes, 
> which is rather a lot.

yes - but one (special) issue here is that there are other 'hot' but 
truly per-CPU structures nearby:

 ffffffff8067e800 D per_cpu__current_kprobe
 ffffffff8067e820 D per_cpu__kprobe_ctlblk
 ffffffff8067e960 D per_cpu__mmu_gathers
 ffffffff8067f960 d per_cpu__runqueues
 ffffffff80680c60 d per_cpu__cpu_domains
 ffffffff80680df0 d per_cpu__sched_group_cpus

cpu_domains is being dirtied too (sd->nr_balance_failed, 
sd->last_balanc, etc.) and mmu_gathers too. So while both mmu_gathers 
and cpu_domains are mostly purely per-CPU, runqueue fields can bounce 
around alot and drag those nearby fields with them (and then get dragged 
back due to those nearby fields being used per-CPU again.)

the runqueue is really supposed to be cacheline-isolated at _both_ ends 
- at its beginning and at its end as well.

> And that putting a gap in the per-cpu memory like this will reduce its 
> overall cache-friendliness.

yes - although the per-cpu runqueue overhead is nearly 5K anyway.

> Remember also that the linesize on VSMP is 4k.

that sucks ...

maybe, to mitigate some of the costs, do a special PER_CPU_CACHE_ALIGNED 
area that collects per-cpu fields that also have significant cross-CPU 
use and need cacheline isolation? Such cacheline-aligned variables, if 
collected separately, would pack up more tightly and would cause only 
half of the wasted space.

	Ingo