Implement K-min-values Sketch for Similarity detection.

This commit is contained in:
Moinak Ghosh 2012-09-11 20:26:36 +05:30
parent 117382c141
commit f3f472b860
4 changed files with 257 additions and 101 deletions

View file

@ -24,8 +24,8 @@
PROG= pcompress
MAINSRCS = main.c utils/utils.c allocator.c zlib_compress.c bzip2_compress.c \
lzma_compress.c ppmd_compress.c adaptive_compress.c lzfx_compress.c \
lz4_compress.c none_compress.c utils/xxhash.c
MAINHDRS = allocator.h pcompress.h utils/utils.h utils/xxhash.h
lz4_compress.c none_compress.c utils/xxhash.c utils/heapq.c
MAINHDRS = allocator.h pcompress.h utils/utils.h utils/xxhash.h utils/heapq.h
MAINOBJS = $(MAINSRCS:.c=.o)
RABINSRCS = rabin/rabin_dedup.c

View file

@ -63,9 +63,12 @@
#include <allocator.h>
#include <utils.h>
#include <pthread.h>
#include <heapq.h>
#include "rabin_dedup.h"
#define FORTY_PCNT(x) (((x)/5 << 1))
extern int lzma_init(void **data, int *level, ssize_t chunksize);
extern int lzma_compress(void *src, size_t srclen, void *dst,
size_t *destlen, int level, uchar_t chdr, void *data);
@ -87,19 +90,7 @@ rabin_min_blksz(uint64_t chunksize, int rab_blk_sz, const char *algo, int delta_
{
uint32_t min_blk;
min_blk = 1 << (rab_blk_sz + RAB_BLK_MIN_BITS);
if (rab_blk_sz > 1)
return (min_blk);
if (((memcmp(algo, "lzma", 4) == 0 || memcmp(algo, "adapt", 5) == 0) &&
chunksize <= LZMA_WINDOW_MAX) || delta_flag) {
if (memcmp(algo, "lzfx", 4) == 0 || memcmp(algo, "lz4", 3) == 0 ||
memcmp(algo, "zlib", 4) == 0 || memcmp(algo, "none", 4) == 0) {
min_blk = 1 << (rab_blk_sz + RAB_BLK_MIN_BITS - 1);
}
} else {
min_blk = 1 << (rab_blk_sz + RAB_BLK_MIN_BITS - 1);
}
return (min_blk);
}
@ -298,32 +289,25 @@ rabin_dedup(rabin_context_t *ctx, uchar_t *buf, ssize_t *size, ssize_t offset, s
char *buf1 = (char *)buf;
uint32_t length;
uint64_t cur_roll_checksum, cur_pos_checksum, cur_sketch;
uint64_t *fplist;
uint32_t len1, fpos[2], cur_sketch2;
uint32_t *charcounts, byts;
uint32_t *fplist;
heap_t heap;
if (rabin_pos == NULL) {
/*
* Initialize arrays for sketch computation. We re-use memory allocated
* for the compressed chunk temporarily.
*/
fplist_sz = 8 * ctx->rabin_poly_avg_block_size;
fplist = (uint64_t *)(ctx->cbuf + ctx->real_chunksize - fplist_sz - 256 * 4);
charcounts = (uint32_t *)(ctx->cbuf + ctx->real_chunksize - 256 * 4);
fplist_sz = 4 * ctx->rabin_poly_max_block_size;
fplist = (uint32_t *)(ctx->cbuf + ctx->real_chunksize - fplist_sz);
memset(fplist, 0, fplist_sz);
memset(charcounts, 0, 256 * 4);
fpos[0] = 0;
fpos[1] = 0;
len1 = 0;
reset_heap(&heap, fplist_sz/2);
}
length = offset;
last_offset = 0;
blknum = 0;
ctx->valid = 0;
cur_roll_checksum = 0;
j = 0;
cur_sketch = 0;
cur_sketch2 = 0;
/*
* If rabin_pos is non-zero then we are being asked to scan for the last rabin boundary
@ -362,6 +346,8 @@ rabin_dedup(rabin_context_t *ctx, uchar_t *buf, ssize_t *size, ssize_t offset, s
}
if (*size < ctx->rabin_poly_avg_block_size) return;
j = 0;
for (i=offset; i<*size; i++) {
uint32_t *splits;
uchar_t cur_byte = buf1[i];
@ -379,68 +365,19 @@ rabin_dedup(rabin_context_t *ctx, uchar_t *buf, ssize_t *size, ssize_t offset, s
cur_pos_checksum = cur_roll_checksum ^ ir[pushed_out];
/*
* Compute a super sketch value of the block. We store a sum of relative
* maximal rabin hash values per 1K(SKETCH_BASIC_BLOCK_SZ) of data. So we
* get upto 128 sums for a max block size of 128K. The bottom blocksize bits
* of the hash are only used which are then biased with the occurrence count.
* This is a representative fingerprint sketch of the block. Storing and
* comparing upto 128 fingerprints per block is very expensive (compute & RAM)
* so we eventually sum all the fingerprints for the block to create a single
* super sketch value representing maximal features of the block. In addition
* the top 2 commonly occuring byte values are used to compute a second sketch
* to refine the earlier one.
* Retain a list of all fingerprints in the block. We then compute
* the K min values sketch from that list and generate a super sketch
* by hashing over the K min values sketch. We only store the least
* significant 32 bits of the fingerprint. This uses less memory,
* requires smaller memset() calls and generates a sufficiently large
* number of similarity matches without false positives - determined
* by experimentation.
*
* This value can be used for similarity detection for delta encoding. Exact
* match for deduplication is additionally detected via a memcmp(). This is a
* variant of some approaches detailed in:
* http://www.armedia.com/wp/SimilarityIndex.pdf
* This is called minhashing and is used widely, for example in various
* search engines to detect similar documents.
*/
len1++;
fpos[1] = cur_pos_checksum & ctx->rabin_avg_block_mask;
splits = (uint32_t *)(&fplist[fpos[1]]);
#if BYTE_ORDER == BIG_ENDIAN
splits[0]++;
splits[1] += cur_pos_checksum & ctx->fp_mask;
#else
splits[1]++;
splits[0] += cur_pos_checksum & ctx->fp_mask;
#endif
charcounts[cur_byte]++;
/*
* Perform the following statement without branching:
* if (fplist[fpos[1]] > fplist[fpos[0]]) fpos[0] = fpos[1];
*/
fpos[0] = fpos[(fplist[fpos[1]] > fplist[fpos[0]])];
if (len1 == SKETCH_BASIC_BLOCK_SZ && ctx->delta_flag) {
uint32_t p1, p2, p3;
/*
* Compute the super sketch value by summing all the representative
* fingerprints of the block.
*/
cur_sketch += fplist[fpos[0]];
memset(fplist, 0, fplist_sz);
fpos[0] = 0;
/*
* Find out the top 2 occurring byte values and compute
* a secondary sketch from them.
*/
p1 = 0;
p2 = 0;
p3 = 0;
for (len1=0; len1<256; len1++) {
if (charcounts[len1] > p1) {
p3 = p2;
p2 = p1;
p1 = len1;
}
charcounts[len1] = 0;
}
cur_sketch2 += ((p1 << 16) | (p2 << 8) | p3);
len1 = 0;
fplist[j] = cur_pos_checksum & 0xFFFFFFFFUL;
j++;
}
/*
* Window pos has to rotate from 0 .. RAB_POLYNOMIAL_WIN_SIZE-1
@ -463,23 +400,26 @@ rabin_dedup(rabin_context_t *ctx, uchar_t *buf, ssize_t *size, ssize_t offset, s
ctx->blocks[blknum]->similar = 0;
ctx->blocks[blknum]->crc = XXH_strong32(buf1+last_offset, length, 0);
// Accumulate the 2 sketch values into a combined similarity checksum
if (ctx->delta_flag) {
ctx->blocks[blknum]->cksum_n_offset = (cur_sketch + cur_sketch2) / 2;
ctx->blocks[blknum]->mean_n_length = cur_sketch / j;
/*
* Reset the heap structure and find the K min values. We use a
* min heap mechanism taken from the heap based priority queue
* implementation in Python.
* Here K = 40%. We are aiming to detect 40% similarity on average.
*/
reset_heap(&heap, FORTY_PCNT(j));
ksmallest(fplist, j, &heap);
cur_sketch = XXH_fast32((const uchar_t *)fplist, FORTY_PCNT(j)*4, 0);
memset(fplist, 0, fplist_sz);
} else {
ctx->blocks[blknum]->cksum_n_offset = 0;
ctx->blocks[blknum]->mean_n_length = 0;
cur_sketch = ctx->blocks[blknum]->crc;
}
fpos[0] = 0;
len1 = 0;
ctx->blocks[blknum]->cksum_n_offset = cur_sketch;
cur_sketch = 0;
blknum++;
last_offset = i+1;
length = 0;
j = 0;
cur_sketch2 = 0;
}
}
@ -500,22 +440,31 @@ rabin_dedup(rabin_context_t *ctx, uchar_t *buf, ssize_t *size, ssize_t offset, s
// Insert the last left-over trailing bytes, if any, into a block.
if (last_offset < *size) {
if (ctx->blocks[blknum] == 0)
ctx->blocks[blknum] = (rabin_blockentry_t *)slab_alloc(NULL, sizeof (rabin_blockentry_t));
ctx->blocks[blknum] = (rabin_blockentry_t *)slab_alloc(NULL,
sizeof (rabin_blockentry_t));
ctx->blocks[blknum]->offset = last_offset;
ctx->blocks[blknum]->index = blknum;
ctx->blocks[blknum]->length = *size - last_offset;
ctx->blocks[blknum]->ref = 0;
ctx->blocks[blknum]->similar = 0;
ctx->blocks[blknum]->crc = XXH_strong32(buf1+last_offset, ctx->blocks[blknum]->length, 0);
if (ctx->delta_flag) {
j = (j > 0 ? j:1);
ctx->blocks[blknum]->cksum_n_offset = (cur_sketch + cur_sketch2) / 2;
ctx->blocks[blknum]->mean_n_length = cur_sketch / j;
if (j > 1) {
reset_heap(&heap, FORTY_PCNT(j));
ksmallest(fplist, j, &heap);
cur_sketch =
XXH_fast32((const uchar_t *)fplist, FORTY_PCNT(j)*4, 0);
} else {
ctx->blocks[blknum]->cksum_n_offset = 0;
ctx->blocks[blknum]->mean_n_length = 0;
cur_sketch =
XXH_fast32((const uchar_t *)fplist, (j*4)/2, 0);
}
ctx->blocks[blknum]->crc = XXH_strong32(buf1+last_offset, ctx->blocks[blknum]->length, 0);
} else {
cur_sketch = ctx->blocks[blknum]->crc;
}
ctx->blocks[blknum]->cksum_n_offset = cur_sketch;
blknum++;
last_offset = *size;
}

193
utils/heapq.c Normal file
View file

@ -0,0 +1,193 @@
/*
* Functions for a rudimentary fast min-heap implementation.
* Derived from Python's _heapqmodule.c by way of drastic simplification
* and a few optimizations.
*/
/*
* Original Python _heapqmodule.c implementation was derived directly
* from heapq.py in Py2.3 which was written by Kevin O'Connor, augmented
* by Tim Peters, annotated by François Pinard, and converted to C by
* Raymond Hettinger.
*/
#include <stdio.h>
#include <limits.h>
#include <stdlib.h>
#include <string.h>
#include <sys/types.h>
#include <stdint.h>
#include <heapq.h>
#ifndef NDEBUG
#define ERROR_CHK
#endif
void
reset_heap(heap_t *heap, __TYPE tot)
{
if (heap) {
heap->len = 0;
heap->tot = tot;
}
}
static int
_siftdownmax(heap_t *h, __TYPE startpos, __TYPE pos)
{
__TYPE newitem, parent;
__TYPE parentpos, *heap;
#ifdef ERROR_CHK
if (pos >= h->len) {
fprintf(stderr, "_siftdownmax: index out of range\n");
return -1;
}
#endif
heap = h->ary;
newitem = heap[pos];
/* Follow the path to the root, moving parents down until finding
a place newitem fits. */
while (pos > startpos){
parentpos = (pos - 1) >> 1;
parent = heap[parentpos];
if (parent < newitem)
break;
heap[pos] = parent;
pos = parentpos;
}
heap[pos] = newitem;
return 0;
}
static int
_siftupmax(heap_t *h, __TYPE spos, __TYPE epos)
{
__TYPE endpos, childpos, rightpos;
__TYPE newitem, *heap, pos;
endpos = h->len;
heap = h->ary;
#ifdef ERROR_CHK
if (pos >= endpos) {
fprintf(stderr, "_siftupmax: index out of range: %u, len: %u\n", pos, endpos);
return -1;
}
#endif
do {
pos = spos;
/* Bubble up the smaller child until hitting a leaf. */
newitem = heap[pos];
childpos = (pos << 1) + 1; /* leftmost child position */
while (childpos < endpos) {
/* Set childpos to index of smaller child. */
rightpos = childpos + 1;
if (rightpos < endpos) {
if (heap[rightpos] < heap[childpos])
childpos = rightpos;
}
/* Move the smaller child up. */
heap[pos] = heap[childpos];
pos = childpos;
childpos = (pos << 1) + 1;
}
/* The leaf at pos is empty now. Put newitem there, and and bubble
it up to its final resting place (by sifting its parents down). */
heap[pos] = newitem;
#ifdef ERROR_CHK
if (_siftdownmax(h, spos, pos) == -1)
return (-1);
#else
_siftdownmax(h, spos, pos);
#endif
spos--;
} while (spos >= epos);
return (0);
}
static int
_siftupmax_s(heap_t *h, __TYPE spos)
{
__TYPE endpos, childpos, rightpos;
__TYPE newitem, *heap, pos;
endpos = h->len;
heap = h->ary;
#ifdef ERROR_CHK
if (pos >= endpos) {
fprintf(stderr, "_siftupmax: index out of range: %u, len: %u\n", pos, endpos);
return -1;
}
#endif
pos = spos;
/* Bubble up the smaller child until hitting a leaf. */
newitem = heap[pos];
childpos = (pos << 1) + 1; /* leftmost child position */
while (childpos < endpos) {
/* Set childpos to index of smaller child. */
rightpos = childpos + 1;
if (rightpos < endpos) {
if (! (heap[rightpos] < heap[childpos]))
childpos = rightpos;
}
/* Move the smaller child up. */
heap[pos] = heap[childpos];
pos = childpos;
childpos = (pos << 1) + 1;
}
/* The leaf at pos is empty now. Put newitem there, and and bubble
it up to its final resting place (by sifting its parents down). */
heap[pos] = newitem;
return (_siftdownmax(h, spos, pos));
}
int
ksmallest(__TYPE *ary, __TYPE len, heap_t *heap)
{
__TYPE elem, los;
__TYPE i, *hp, n;
#ifdef ERROR_CHK
if (len >= heap->tot) {
fprintf(stderr, "nsmallest: array size > heap size\n");
return (-1);
}
#endif
n = heap->tot;
heap->ary = ary;
hp = ary;
heap->len = n;
#ifdef ERROR_CHK
if(_siftupmax(heap, n/2-1, 0) == -1)
return (-1);
#else
_siftupmax(heap, n/2-1, 0);
#endif
los = hp[0];
for (i = n; i < len; i++) {
elem = ary[i];
if (elem >= los) {
continue;
}
hp[0] = elem;
#ifdef ERROR_CHK
if (_siftupmax_s(heap, 0) == -1)
return (-1);
#else
_siftupmax_s(heap, 0);
#endif
los = hp[0];
}
return 0;
}

14
utils/heapq.h Normal file
View file

@ -0,0 +1,14 @@
#ifndef __HEAPQ_H_
#define __TYPE int32_t
typedef struct {
__TYPE *ary;
__TYPE len;
__TYPE tot;
} heap_t;
extern int ksmallest(__TYPE *ary, __TYPE len, heap_t *heap);
extern void reset_heap(heap_t *h, __TYPE tot);
#endif