[dpdk-dev] [PATCH v3 1/4] hash: reorder hash structure

De Lara Guarch, Pablo pablo.de.lara.guarch at intel.com
Thu Sep 29 03:33:12 CEST 2016



> -----Original Message-----
> From: Richardson, Bruce
> Sent: Wednesday, September 28, 2016 2:03 AM
> To: De Lara Guarch, Pablo
> Cc: dev at dpdk.org
> Subject: Re: [PATCH v3 1/4] hash: reorder hash structure
> 
> On Tue, Sep 06, 2016 at 08:34:01PM +0100, Pablo de Lara wrote:
> > In order to optimize lookup performance, hash structure
> > is reordered, so all fields used for lookup will be
> > in the first cache line.
> >
> > Signed-off-by: Pablo de Lara <pablo.de.lara.guarch at intel.com>
> > ---
> >  lib/librte_hash/rte_cuckoo_hash.h | 12 +++++++-----
> >  1 file changed, 7 insertions(+), 5 deletions(-)
> >
> > diff --git a/lib/librte_hash/rte_cuckoo_hash.h
> b/lib/librte_hash/rte_cuckoo_hash.h
> > index e290dab..701531a 100644
> > --- a/lib/librte_hash/rte_cuckoo_hash.h
> > +++ b/lib/librte_hash/rte_cuckoo_hash.h
> > @@ -182,9 +182,7 @@ struct rte_hash_bucket {
> >
> >  /** A hash table structure. */
> >  struct rte_hash {
> > -	char name[RTE_HASH_NAMESIZE];   /**< Name of the hash. */
> > -	uint32_t entries;               /**< Total table entries. */
> > -	uint32_t num_buckets;           /**< Number of buckets in table. */
> > +	/* first cache line - fields used in lookup */
> >  	uint32_t key_len;               /**< Length of hash key. */
> >  	rte_hash_function hash_func;    /**< Function used to calculate hash.
> */
> >  	uint32_t hash_func_init_val;    /**< Init value used by hash_func. */
> > @@ -196,12 +194,13 @@ struct rte_hash {
> >  						from hash signature. */
> >  	uint32_t key_entry_size;         /**< Size of each key entry. */
> >
> > -	struct rte_ring *free_slots;    /**< Ring that stores all indexes
> > -						of the free slots in the key
> table */
> >  	void *key_store;                /**< Table storing all keys and data */
> >  	struct rte_hash_bucket *buckets;	/**< Table with buckets
> storing all the
> >  							hash values and key
> indexes
> >  							to the key table*/
> > +
> > +	struct rte_ring *free_slots;    /**< Ring that stores all indexes
> > +						of the free slots in the key
> table */
> >  	uint8_t hw_trans_mem_support;	/**< Hardware transactional
> >  							memory support */
> >  	struct lcore_cache *local_free_slots;
> > @@ -209,6 +208,9 @@ struct rte_hash {
> >  	enum add_key_case add_key; /**< Multi-writer hash add behavior */
> >
> >  	rte_spinlock_t *multiwriter_lock; /**< Multi-writer spinlock for w/o
> TM */
> > +	char name[RTE_HASH_NAMESIZE];   /**< Name of the hash. */
> > +	uint32_t entries;               /**< Total table entries. */
> > +	uint32_t num_buckets;           /**< Number of buckets in table. */
> 
> Hi Pablo,
> 
> While I've no strong objection to the change, having the name at the start
> is a common paradigm in DPDK. Rather than place these fields at the end,
> can you get the same effect by just marking the key_len function
> __rte_cache_aligned? It may use a little more memory per table, but given
> that
> the size of the hash table is going to be largely governed by the table data,
> I don't see an extra 64 bytes in the structure as being an issue.

Hi Bruce,

Sure, sounds good to me. I was trying a similar approach to the mbuf structure,
and actually I saw a small performance boost doing this,
so I can push the fields needed for lookup for the second cache line instead,
it should work the same :)

Thanks,
Pablo
> 
> Regards,
> /Bruce


More information about the dev mailing list