Re: [PATCH 4/3] Replace dynamic percpu implementation

[Ravikiran G Thirumalai <kiran@in.ibm.com>]

On Thu, May 22, 2003 at 06:36:31PM +1000, Rusty Russell wrote:
> In message <20030522081423.GC27614@in.ibm.com> you write:
> > 4. Extra dereferences in alloc_percpu were not significant, but alloc_percpu
> >    was interlaced and kmalloc_percpu_new wasn't.  Insn profile seemed to
> >    indicate extra cost in memory dereferencing of alloc_percpu was
> >    offset by the interlacing/objects sharing the same cacheline part.
> >    but then insn profiles are only indicative...not accurate.
> 
> Interesting: personally I consider the cacheline sharing a feature,
> and unless you've done something special, the static declaration
> should be interlaced too, no?

Yes, the static declartion was interlaced too. What I meant to say is that
cacheline sharing feature helped alloc_percpu/static percpu, compensate
for the small extra memory reference cost in getting __percpu_offset[]
when you compare with kmalloc_percpu_new.

>... 
> Aside: if kmalloc_percpu uses the per-cpu offset too, it probably
> makes sense to make the per-cpu offset to a first class citizen, and
> smp_processor_id to be derived, rather than the other way around as at
> the moment.  This would offer further speedup by removing a level of
> indirection.
> 
> If you're interested I can probably produce such a patch for x86...

Sure, it might help per-cpu data but will it cause performance
regression elsewhere? (other users of smp_processor_id).  I can run it 
through the same tests and find out.  Maybe it'll make good paper material 
for later? ;)


Thanks,
Kiran