Re: [PATCH] libnuma: disable caching of node cpusmasks

[Petr Holasek <pholasek@redhat.com>]

On Fri, 12 Jun 2015, Andi Kleen <andi@firstfloor.org> wrote:
> > @@ -1329,25 +1321,12 @@ numa_node_to_cpus_v2(int node, struct bitmask *buffer)
> >  	size_t len = 0; 
> >  	struct bitmask *mask;
> >  
> > -	if (!node_cpu_mask_v2)
> > -		init_node_cpu_mask_v2();
> > -
> >  	if (node > nnodes) {
> >  		errno = ERANGE;
> >  		return -1;
> >  	}
> >  	numa_bitmask_clearall(buffer);
> >  
> > -	if (node_cpu_mask_v2[node]) {
> 
> I was playing around with info (http://fbinfer.com/) earlier, and ran it over
> numactl, and it complained about exactly this line: it can reference NULL
> when the memory allocation above fails. So it's good to remove it.
> 
> Would be good to see how much performance difference it makes though.
> Do you have any data? If it's significant may need to do a time out
> or similar.
> 

I did simple comparison for 1 million of numa_node_to_cpus() calls
and disabled caching extended time three times (from 10 seconds to 30).

-- 
Petr Holasek
pholasek@redhat.com