Re: [PATCH] net: fold network name hash (v2)

[Stephen Hemminger <shemminger@vyatta.com>]

On Wed, 28 Oct 2009 07:07:10 +0100
Eric Dumazet <eric.dumazet@gmail.com> wrote:

> Stephen Hemminger a écrit :
> > The full_name_hash does not produce a value that is evenly distributed
> > over the lower 8 bits. This causes name hash to be unbalanced with large
> > number of names. There is a standard function to fold in upper bits
> > so use that.
> > 
> > This is independent of possible improvements to full_name_hash()
> > in future.
> 
> >  static inline struct hlist_head *dev_name_hash(struct net *net, const char *name)
> >  {
> >  	unsigned hash = full_name_hash(name, strnlen(name, IFNAMSIZ));
> > -	return &net->dev_name_head[hash & ((1 << NETDEV_HASHBITS) - 1)];
> > +	return &net->dev_name_head[hash_long(hash, NETDEV_HASHBITS)];
> >  }
> >  
> >  static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex)
> 
> full_name_hash() returns an "unsigned int", which is guaranteed to be 32 bits
> 
> You should therefore use hash_32(hash, NETDEV_HASHBITS),
> not hash_long() that maps to hash_64() on 64 bit arches, which is
> slower and certainly not any better with a 32bits input.

OK, I was following precedent. Only a couple places use hash_32, most use
hash_long().

Using the upper bits does give better distribution.
With 100,000 network names:

               Time       Ratio       Max   StdDev
hash_32       0.002123     1.00       422  11.07
hash_64       0.002927     1.00       400   3.97

The time field is pretty meaningless for such a small sample