[dpdk-dev] [PATCH v4 3/7] member: implement vBF mode

Wang, Yipeng1 yipeng1.wang at intel.com
Tue Oct 3 03:24:38 CEST 2017


Please see one comment inlined:

> -----Original Message-----
> From: De Lara Guarch, Pablo
> Sent: Monday, October 2, 2017 8:44 AM
> To: Wang, Yipeng1 <yipeng1.wang at intel.com>; dev at dpdk.org
> Cc: thomas at monjalon.net; Tai, Charlie <charlie.tai at intel.com>; Gobriel,
> Sameh <sameh.gobriel at intel.com>; Mcnamara, John
> <john.mcnamara at intel.com>
> Subject: RE: [PATCH v4 3/7] member: implement vBF mode
> 
> 
> 
> > -----Original Message-----
> > From: Wang, Yipeng1
> > Sent: Wednesday, September 27, 2017 6:41 PM
> > To: dev at dpdk.org
> > Cc: thomas at monjalon.net; Tai, Charlie <charlie.tai at intel.com>; Gobriel,
> > Sameh <sameh.gobriel at intel.com>; De Lara Guarch, Pablo
> > <pablo.de.lara.guarch at intel.com>; Mcnamara, John
> > <john.mcnamara at intel.com>; Wang, Yipeng1 <yipeng1.wang at intel.com>
> > Subject: [PATCH v4 3/7] member: implement vBF mode
> >
> > Bloom Filter (BF) [1] is a well-known space-efficient probabilistic data
> > structure that answers set membership queries.
> > Vector of Bloom Filters (vBF) is an extension to traditional BF that supports
> > multi-set membership testing. Traditional BF will return found or not-found
> > for each key. vBF will also return which set the key belongs to if it is found.
> >
> > Since each set requires a BF, vBF should be used when set count is small.
> > vBF's false positive rate could be set appropriately so that its memory
> > requirement and lookup speed is better in certain cases comparing to HT
> > based set-summary.
> >
> > This patch adds the vBF implementation.
> >
> > [1]B H Bloom, “Space/Time Trade-offs in Hash Coding with Allowable
> > Errors,” Communications of the ACM, 1970.
> >
> > Signed-off-by: Yipeng Wang <yipeng1.wang at intel.com>
> 
> ...
> 
> > diff --git a/lib/librte_member/rte_member_vbf.c
> > b/lib/librte_member/rte_member_vbf.c
> 
> ...
> 
> > +int
> > +rte_member_create_vbf(struct rte_member_setsum *ss,
> > +		const struct rte_member_parameters *params) {
> > +
> > +	if (params->num_set > 32 || !rte_is_power_of_2(params-
> > >num_set) ||
> 
> Magic number. Define a macro instead.
> 
> > +			params->num_keys == 0 ||
> > +			params->false_positive_rate == 0 ||
> > +			params->false_positive_rate > 1) {
> > +		rte_errno = EINVAL;
> > +		RTE_MEMBER_LOG(ERR, "vBF create with invalid
> > parameters\n");
> > +		return -EINVAL;
> 
> ...
> 
> > +
> > +	/*
> > +	 * reduce hash function count, until we approach the user specified
> > +	 * false-positive rate. otherwise it is too conservative
> 
> Watch out for capital letters at the start of the comment and after a full stop.
> 
> > +	 */
> > +	int tmp_num_hash = ss->num_hashes;
> > +
> > +	while (tmp_num_hash > 1) {
> > +		float tmp_fp = new_fp;
> > +
> > +		tmp_num_hash--;
> > +		new_fp = pow((1 - pow((1 - 1.0 / ss->bits),
> > num_keys_per_bf *
> > +					tmp_num_hash)), tmp_num_hash);
> > +		new_fp = 1 - pow((1 - new_fp), ss->num_set);
> > +
> > +		if (new_fp > params->false_positive_rate) {
> > +			new_fp = tmp_fp;
> > +			tmp_num_hash++;
> > +			break;
> > +		}
> > +	}
> > +
> > +	ss->num_hashes = tmp_num_hash;
> > +
> > +	RTE_MEMBER_LOG(DEBUG, "vector bloom filter created, "
> > +		"each bloom filter expects %u keys, needs %u bits, %u
> > hashes, "
> > +		"with false positive rate set as %.5f, "
> > +		"The new calculated vBF false positive rate is %.5f\n",
> > +		num_keys_per_bf, ss->bits, ss->num_hashes, x, new_fp);
> 
> Use a more descriptive variable name for "x".
> 
> > +
> > +	ss->table = rte_zmalloc_socket(NULL, ss->num_set * (ss->bits >> 3),
> > +					RTE_CACHE_LINE_SIZE, ss-
> > >socket_id);
> > +
> > +	/*
> > +	 * To avoid multiplication and division:
> > +	 * mul_shift is used for multiplication shift during bit test
> > +	 * div_shift is used for division shift, to be divided by number of bits
> > +	 * represented by a uint32_t variable
> > +	 */
> > +	ss->mul_shift = __builtin_ctzl(ss->num_set);
> > +	ss->div_shift = __builtin_ctzl(32 >> ss->mul_shift);
> > +
> > +	if (ss->table == NULL)
> > +		return -ENOMEM;
> 
> I would move this check just after the malloc call.
> 
> > +
> > +	return 0;
> > +}
> > +
> > +static inline uint32_t
> > +test_bit(uint32_t bit_loc, const struct rte_member_setsum *ss) {
> > +	uint32_t *vbf = ss->table;
> > +	uint32_t n = ss->num_set;
> > +	uint32_t div_shift = ss->div_shift;
> > +	uint32_t mul_shift = ss->mul_shift;
> > +	/*
> > +	 * a is how many bits in one BF are represented by one 32bit
> > +	 * variable.
> > +	 */
> > +	uint32_t a = 32 >> mul_shift;
> > +	/*
> > +	 * x>>b is the divide, x & (a-1) is the mod, & (1<<n-1) to mask out
> > bits
> > +	 * we do not need
> > +	 */
> > +	return (vbf[bit_loc>>div_shift] >> ((bit_loc & (a - 1)) << mul_shift))
> 
> Add spaces around ">>".
> 
> > &
> > +							((1ULL << n) - 1);
> > +}
> > +
> > +static inline void
> > +set_bit(uint32_t bit_loc, const struct rte_member_setsum *ss, int32_t
> > +set) {
> > +	uint32_t *vbf = ss->table;
> > +	uint32_t div_shift = ss->div_shift;
> > +	uint32_t mul_shift = ss->mul_shift;
> > +	uint32_t a = 32 >> mul_shift;
> > +
> > +	vbf[bit_loc>>div_shift] |= 1U << (((bit_loc & (a - 1)) << mul_shift) +
> > +								set - 1);
> 
> Same as above.
> 
> > +}
> > +
> > +int
> > +rte_member_lookup_vbf(const struct rte_member_setsum *ss, const
> > void *key,
> > +		member_set_t *set_id)
> > +{
> > +	uint32_t j;
> > +	uint32_t h1 = MEMBER_HASH_FUNC(key, ss->key_len, ss-
> > >prim_hash_seed);
> > +	uint32_t h2 = MEMBER_HASH_FUNC(&h1, sizeof(uint32_t),
> > +						ss->sec_hash_seed);
> > +	uint32_t mask = ~0;
> > +	uint32_t bit_loc;
> > +
> > +	for (j = 0; j < ss->num_hashes; j++) {
> > +		bit_loc = (h1 + j * h2) & ss->bit_mask;
> > +		mask &= test_bit(bit_loc, ss);
> > +	}
> > +
> > +	if (mask) {
> > +		*set_id = __builtin_ctzl(mask) + 1;
> > +		return 1;
> 
> Wouldn't it be better to return 0 when there is a hit and -ENOENT when
> there is not?
[Wang, Yipeng] 
I tried to follow the convention of other lookup functions that the return value
Is kind of the number of matches that found. So 1 means found 1 and 0 means not
Found. I defined the public API rte_member_lookup to also returns 1 for found.

> > +	}
> > +
> > +	*set_id = RTE_MEMBER_NO_MATCH;
> > +	return 0;
> > +}
> > +
> > +uint32_t
> > +rte_member_lookup_bulk_vbf(const struct rte_member_setsum *ss,
> > +		const void **keys, uint32_t num_keys, member_set_t
> > *set_ids) {
> > +	uint32_t i, k;
> > +	uint32_t ret = 0;
> 
> Change variable name to "nr_matches/hits" or similar.
> Same in the next functions.


More information about the dev mailing list