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

[Andi Kleen <ak@suse.de>]

> >  
> > -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, 

On x86 just the real possible map now -- that tends to be much smaller.

There might be some other architectures who still allocate per cpu
for all of NR_CPUs (or always set possible map to that), but those
should be just fixed.

> which is 
> rather a lot.

We should have solved the problem of limited per cpu space in .22 at least
with some patches by Jeremy. I also plan a few other changes the will
use more per CPU memory again.

> Remember also that the linesize on VSMP is 4k.
> 
> And that putting a gap in the per-cpu memory like this will reduce its
> overall cache-friendliness.

When he avoids false sharing on remote wakeup it should be more cache friendly.

> Need more convincing, please.

Was this based on some benchmark where it showed?

-Andi