Re: [PATCH v2] - Randomize node rotor used in cpuset_mem_spread_node()

[Matt Mackall <mpm@selenic.com>]

On Wed, 2010-04-28 at 15:40 -0700, Andrew Morton wrote:
> On Wed, 28 Apr 2010 10:04:32 -0500
> Jack Steiner <steiner@sgi.com> wrote:
> 
> > Some workloads that create a large number of small files tend to assign
> > too many pages to node 0 (multi-node systems). Part of the reason is that
> > the rotor (in cpuset_mem_spread_node()) used to assign nodes starts
> > at node 0 for newly created tasks.
> 
> And, presumably, your secret testcase forks lots of subprocesses which
> do the file creation?
> 
> > This patch changes the rotor to be initialized to a random node number
> > of the cpuset.
> 
> Why random as opposed to, say, inherit-rotor-from-parent?

That'd be fine, I bet.

> > Index: linux/arch/x86/mm/numa.c
> > ===================================================================
> > --- linux.orig/arch/x86/mm/numa.c	2010-04-28 09:44:52.422898844 -0500
> > +++ linux/arch/x86/mm/numa.c	2010-04-28 09:49:39.282899779 -0500
> > @@ -2,6 +2,7 @@
> >  #include <linux/topology.h>
> >  #include <linux/module.h>
> >  #include <linux/bootmem.h>
> > +#include <linux/random.h>
> >  
> >  #ifdef CONFIG_DEBUG_PER_CPU_MAPS
> >  # define DBG(x...) printk(KERN_DEBUG x)
> > @@ -65,3 +66,19 @@ const struct cpumask *cpumask_of_node(in
> >  }
> >  EXPORT_SYMBOL(cpumask_of_node);
> >  #endif
> > +
> > +/*
> > + * Return the bit number of a random bit set in the nodemask.
> > + *   (returns -1 if nodemask is empty)
> > + */
> > +int __node_random(const nodemask_t *maskp)
> > +{
> > +	int w, bit = -1;
> > +
> > +	w = nodes_weight(*maskp);
> > +	if (w)
> > +		bit = bitmap_find_nth_bit(maskp->bits,
> > +			get_random_int() % w, MAX_NUMNODES);
> > +	return bit;
> > +}
> > +EXPORT_SYMBOL(__node_random);
> 
> I suspect random32() would suffice here.  It avoids depleting the
> entropy pool altogether.

I wouldn't worry about that. get_random_int() touches the urandom pool,
which will always leave entropy around. Also, Ted and I decided over a
year ago that we should drop the whole entropy accounting framework,
which I'll get around to some rainy weekend.

-- 
http://selenic.com : development and support for Mercurial and Linux