agora inbox for [email protected]  
help / color / mirror / Atom feed
From: Tomas Vondra <[email protected]>
Subject: [PATCH 6/8] use one-hash bloom variant
Date: Mon, 2 Nov 2020 23:55:28 +0100

one-hash tweaks

tweak generation of primes
---
 src/backend/access/brin/brin_bloom.c | 179 ++++++++++++++++++++++++---
 1 file changed, 159 insertions(+), 20 deletions(-)

diff --git a/src/backend/access/brin/brin_bloom.c b/src/backend/access/brin/brin_bloom.c
index f7b405f76f..87a3d5598a 100644
--- a/src/backend/access/brin/brin_bloom.c
+++ b/src/backend/access/brin/brin_bloom.c
@@ -208,6 +208,9 @@ typedef struct BloomOptions
 #define		BLOOM_DEFAULT_FALSE_POSITIVE_RATE	0.01	/* 1% fp rate */
 #define		BLOOM_DEFAULT_SORT_MODE			true		/* start in sort */
 
+/* With the minimum allowed false positive rate of 0.001, we need up to 10 hashes */
+#define		BLOOM_MAX_NUM_PARTITIONS	10
+
 #define BloomGetNDistinctPerRange(opts) \
 	((opts) && (((BloomOptions *) (opts))->nDistinctPerRange != 0) ? \
 	 (((BloomOptions *) (opts))->nDistinctPerRange) : \
@@ -279,6 +282,7 @@ typedef struct BloomFilter
 	uint8	nhashes;	/* number of hash functions */
 	uint32	nbits;		/* number of bits in the bitmap (size) */
 	uint32	nbits_set;	/* number of bits set to 1 */
+	uint32	partlens[BLOOM_MAX_NUM_PARTITIONS];	/* partition lengths */
 
 	/* data of the bloom filter (used both for sorted and hashed phase) */
 	char	data[FLEXIBLE_ARRAY_MEMBER];
@@ -288,6 +292,127 @@ typedef struct BloomFilter
 static BloomFilter *bloom_switch_to_hashing(BloomFilter *filter);
 
 
+/*
+ * generate_primes
+ * 		returns array of all primes less than limit
+ *
+ * WIP: very naive prime sieve; could be optimized using segmented ranges
+ */
+static uint32 *
+generate_primes(int limit)
+{
+	/* upper bound of number of primes below limit */
+	/* WIP: reference for this number */
+	int numprimes = 1.26 * limit / log(limit);
+
+	bool *is_composite = (bool *) palloc0(limit * sizeof(bool));
+	uint32 *primes = (uint32 *) palloc0(numprimes * sizeof(uint32));
+
+	int maxfactor = floor(sqrt(limit));
+	int factor = 2;	/* first prime */
+
+	/* mark the sieve where the index is composite */
+	while (factor < maxfactor)
+	{
+		for (int i = factor * factor; i < limit; i += factor)
+			 is_composite[i] = true;
+		do { factor++; } while (is_composite[factor]);
+	}
+
+	/* the unmarked numbers are prime, so copy over */
+	for (int i = 2, j = 0; i < limit && j < numprimes; i++)
+	{
+		if (!is_composite[i])
+			primes[j++] = i;
+	}
+
+	/* there should still be some zeroes at the end, but make sure */
+	primes[numprimes - 1] = 0;
+
+	/* pretty large, so free it now (segmented ranges would make it smaller) */
+	pfree(is_composite);
+	return primes;
+}
+
+/*
+ * set_bloom_partitions
+ * 		Calculate k moduli for one-hashing bloom filter.
+ *
+ * Find consecutive primes whose sum is close to nbits and
+ * return the sum. Copy the primes to the filter to use as
+ * partition lengths.
+ * WIP: one-hashing bf paper ref somewhere
+ */
+static uint32
+set_bloom_partitions(int nbits, uint8 nhashes, uint32 *partlens)
+{
+	int		min, diff, incr;
+	int		pidx = 0;
+	int		sum = 0;
+
+	/* we want partitions roughly with this length */
+	int		target_partlen = nbits / nhashes;
+	uint32 *primes;
+
+	/*
+	 * Generate primes up to a maximum value, based on the target length.
+	 * How much higher it needs to be is based on gaps between primes,
+	 * as shown at: https://primes.utm.edu/notes/gaps.html
+	 *
+	 * The largest possible filter is 32kB (the largest page size), i.e.
+	 * ~262kb. Assuming a single partition of this length, the maximum
+	 * gap is 95. With more partitions the gaps would be smaller, so we
+	 * just use this as an upper boundary, and multiply it with the max
+	 * number of partitions.
+	 *
+	 * Note: This is an overkill, but the impact on CPU time is minimal,
+	 * particularly with "regular" filter sizes. If needed, we can make
+	 * this more efficient in the future.
+	 */
+	primes = generate_primes(target_partlen + BLOOM_MAX_NUM_PARTITIONS * 100);
+
+	/*
+	 * In our array of primes, find a sequence of length nhashes, whose
+	 * last item is close to our target partition length. The end of the
+	 * array will be filled with zeros, so we need to guard against that.
+	 */
+	while (primes[pidx + nhashes - 1] <= target_partlen &&
+		   primes[pidx + nhashes] > 0)
+		pidx++;
+
+	for (int i = 0; i < nhashes; i++)
+		sum += primes[pidx + i];
+
+	/*
+	 * Since all the primes are less than or equal the desired partition
+	 * length, the sum is somewhat less than nbits. Increment the starting
+	 * point until we find the sequence of primes whose sum is closest to
+	 * nbits. It doesn't matter whether it's higher or lower.
+	 */
+	min = abs(nbits - sum);
+	for (;;)
+	{
+		incr = primes[pidx + nhashes] - primes[pidx];
+		diff = abs(nbits - (sum + incr));
+		if (diff >= min)
+			break;
+
+		min = diff;
+		sum += incr;
+		pidx++;
+	}
+
+	memcpy(partlens, &primes[pidx], nhashes * sizeof(uint32));
+
+	/* WIP: assuming it's not important to pfree primes */
+
+	/*
+	 * The actual filter length will be the sum of the partition lengths
+	 * rounded up to the nearest byte.
+	 */
+	return (uint32) ((sum + 7) / 8) * 8;
+}
+
 /*
  * bloom_init
  * 		Initialize the Bloom Filter, allocate all the memory.
@@ -319,6 +444,7 @@ bloom_init(bool sort_mode, int ndistinct, double false_positive_rate)
 	 */
 	k = log(2.0) * m / ndistinct;
 	k = (k - floor(k) >= 0.5) ? ceil(k) : floor(k);
+	k = Min(k, BLOOM_MAX_NUM_PARTITIONS);
 
 	/*
 	 * When sort phase is enabled, start with a small filter which we grow
@@ -338,7 +464,9 @@ bloom_init(bool sort_mode, int ndistinct, double false_positive_rate)
 
 	filter->flags = 0;
 	filter->nhashes = (int) k;
-	filter->nbits = m;
+
+	/* calculate the partition lengths and adjust m to match */
+	filter->nbits = set_bloom_partitions(m, k, filter->partlens);
 
 	if (!sort_mode)
 		filter->flags |= BLOOM_FLAG_PHASE_HASH;
@@ -463,7 +591,7 @@ static BloomFilter *
 bloom_add_value(BloomFilter *filter, uint32 value, bool *updated)
 {
 	int		i;
-	uint32	h1, h2;
+	int		part_boundary = 0;
 
 	/* assume 'not updated' by default */
 	Assert(filter);
@@ -532,17 +660,16 @@ bloom_add_value(BloomFilter *filter, uint32 value, bool *updated)
 	/* we better be in the hashing phase */
 	Assert(BLOOM_IS_HASHED(filter));
 
-	/* compute the hashes, used for the bloom filter */
-	h1 = hash_uint32_extended(value, 0x71d924af) % filter->nbits;
-	h2 = hash_uint32_extended(value, 0xba48b314) % filter->nbits;
-
 	/* compute the requested number of hashes */
 	for (i = 0; i < filter->nhashes; i++)
 	{
-		/* h1 + h2 + f(i) */
-		uint32	h = (h1 + i * h2) % filter->nbits;
-		uint32	byte = (h / 8);
-		uint32	bit  = (h % 8);
+		int partlen = filter->partlens[i];
+		int bitloc = part_boundary + (value % partlen);
+
+		int byte = (bitloc / 8);
+		int bit  = (bitloc % 8);
+
+		Assert(bitloc < filter->nbits);
 
 		/* if the bit is not set, set it and remember we did that */
 		if (! (filter->data[byte] & (0x01 << bit)))
@@ -552,6 +679,9 @@ bloom_add_value(BloomFilter *filter, uint32 value, bool *updated)
 			if (updated)
 				*updated = true;
 		}
+
+		/* next bit */
+		part_boundary += partlen;
 	}
 
 	return filter;
@@ -581,6 +711,7 @@ bloom_switch_to_hashing(BloomFilter *filter)
 
 	newfilter->nhashes = filter->nhashes;
 	newfilter->nbits = filter->nbits;
+	memcpy(newfilter->partlens, filter->partlens, filter->nhashes * sizeof(uint32));
 	newfilter->flags |= BLOOM_FLAG_PHASE_HASH;
 
 	SET_VARSIZE(newfilter, len);
@@ -605,7 +736,7 @@ static bool
 bloom_contains_value(BloomFilter *filter, uint32 value)
 {
 	int		i;
-	uint32	h1, h2;
+	int		part_boundary = 0;
 
 	Assert(filter);
 
@@ -634,21 +765,23 @@ bloom_contains_value(BloomFilter *filter, uint32 value)
 	/* now the regular hashing mode */
 	Assert(BLOOM_IS_HASHED(filter));
 
-	/* calculate the two hashes */
-	h1 = hash_uint32_extended(value, 0x71d924af) % filter->nbits;
-	h2 = hash_uint32_extended(value, 0xba48b314) % filter->nbits;
-
 	/* compute the requested number of hashes */
 	for (i = 0; i < filter->nhashes; i++)
 	{
-		/* h1 + h2 + f(i) */
-		uint32	h = (h1 + i * h2) % filter->nbits;
-		uint32	byte = (h / 8);
-		uint32	bit  = (h % 8);
+		int partlen = filter->partlens[i];
+		int bitloc = part_boundary + (value % partlen);
+
+		int byte = (bitloc / 8);
+		int bit  = (bitloc % 8);
+
+		Assert(bitloc < filter->nbits);
 
 		/* if the bit is not set, the value is not there */
 		if (! (filter->data[byte] & (0x01 << bit)))
 			return false;
+
+		/* next bit */
+		part_boundary += partlen;
 	}
 
 	/* all hashes found in bloom filter */
@@ -1081,8 +1214,14 @@ brin_bloom_summary_out(PG_FUNCTION_ARGS)
 
 	if (BLOOM_IS_HASHED(filter))
 	{
-		appendStringInfo(&str, "mode: hashed  nhashes: %u  nbits: %u  nbits_set: %u",
+		appendStringInfo(&str,
+						 "mode: hashed  nhashes: %u  nbits: %u  nbits_set: %u  partition lengths:  [",
 						 filter->nhashes, filter->nbits, filter->nbits_set);
+		for (int i = 0; i < filter->nhashes - 1; i++)
+		{
+			appendStringInfo(&str, "%u, ", filter->partlens[i]);
+		}
+		appendStringInfo(&str, "%u]", filter->partlens[filter->nhashes - 1]);
 	}
 	else
 	{
-- 
2.26.2


--------------45196B023835614BDFCBD16D
Content-Type: text/x-patch; charset=UTF-8;
 name="0005-use-two-independent-hashes-20201220.patch"
Content-Transfer-Encoding: 7bit
Content-Disposition: attachment;
 filename="0005-use-two-independent-hashes-20201220.patch"



view thread (5+ messages)  latest in thread

reply

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Reply to all the recipients using the --to and --cc options:
  reply via email

  To: [email protected]
  Cc: [email protected]
  Subject: Re: [PATCH 6/8] use one-hash bloom variant
  In-Reply-To: <no-message-id-175373@localhost>

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox