Speed up adaptive modes by using heuristics to select compression algorithm.

Select similarity percentage based on dedupe block size for effectiveness.

README.md

@@ -82,11 +82,13 @@ NOTE: The option "libbsc" uses  Ilya Grebnov's block sorting compression library
 
     Perform Delta Encoding in addition to Identical Dedup:
        pcompress -E ... - This also implies '-D'. This performs Delta Compression
-                          between 2 blocks if they are at least 60% similar.
+                          between 2 blocks if they are 40% to 60% similar. The
+                          similarity %age is selected based on the dedupe block
+                          size to balance performance and effectiveness.
        pcompress -EE .. - This causes Delta Compression to happen if 2 blocks are
-                          at least 40% similar. This can effect greater final
+                          at least 40% similar regardless of block size. This can
-                          compression ratio at the cost of higher processing
+                          effect greater final compression ratio at the cost of
-                          overhead.
+                          higher processing overhead.
 
     Number of threads can optionally be specified: -t <1 - 256 count>
     Other flags:

adaptive_compress.c

@@ -140,55 +140,43 @@ adapt_compress(void *src, size_t srclen, void *dst,
 	size_t *dstlen, int level, uchar_t chdr, void *data)
 {
 	struct adapt_data *adat = (struct adapt_data *)(data);
-	int rv, rv1, rv2;
+	uchar_t *src1 = (uchar_t *)src;
-	unsigned int *inc;
+	size_t i, bincount;
-	size_t dst2len, dst3len, smaller_dstlen;
+	int rv;
-	uchar_t *dst2, *smaller_dst;
-	void *tmp;
 
-	dst2 = slab_alloc(NULL, *dstlen);
+	/*
-	if (!dst2) {
+	 * Count number of 8-bit binary bytes in source.
-		fprintf(stderr, "Adapt: Out of memory\n");
+	 */
-		return (-1);
+	bincount = 0;
-	}
+	for (i = 0; i < srclen; i++)
+		bincount += (src1[i] >> 7);
 
-	rv = COMPRESS_PPMD;
+	/*
-	inc = &ppmd_count;
+	 * Use PPMd if at least 70% of source is 7-bit textual bytes, otherwise
-	dst2len = *dstlen;
+	 * use Bzip2 or LZMA.
-	dst3len = *dstlen;
+	 */
-	rv1 = ppmd_compress(src, srclen, dst, dstlen, level, chdr, adat->ppmd_data);
+	if (bincount > (srclen / 10 * 3)) {
-	if (rv1 < 0) *dstlen = dst3len;
+		if (adat->adapt_mode == 2) {
-
+			rv = lzma_compress(src, srclen, dst, dstlen, level, chdr, adat->lzma_data);
-	if (adat->adapt_mode == 2) {
+			if (rv < 0)
-		rv2 = lzma_compress(src, srclen, dst2, &dst2len, level, chdr, adat->lzma_data);
+				return (rv);
-		if (rv2 < 0) dst2len = dst3len;
-		if (dst2len < *dstlen) {
-			inc = &lzma_count;
 			rv = COMPRESS_LZMA;
-		}
+			lzma_count++;
-	} else {
+		} else {
-		rv2 = bzip2_compress(src, srclen, dst2, &dst2len, level, chdr, NULL);
+			rv = bzip2_compress(src, srclen, dst, dstlen, level, chdr, NULL);
-		if (rv2 < 0) dst2len = dst3len;
+			if (rv < 0)
-		if (dst2len < *dstlen) {
+				return (rv);
-			inc = &bzip2_count;
 			rv = COMPRESS_BZIP2;
+			bzip2_count++;
 		}
-	}
-
-	if (dst2len < *dstlen) {
-		smaller_dstlen = dst2len;
-		smaller_dst = dst2;
 	} else {
-		smaller_dstlen = *dstlen;
+		rv = ppmd_compress(src, srclen, dst, dstlen, level, chdr, adat->ppmd_data);
-		smaller_dst = dst;
+		if (rv < 0)
+			return (rv);
+		rv = COMPRESS_PPMD;
+		ppmd_count++;
 	}
 
-	*inc += 1;
-	if (smaller_dst != dst) {
-		memcpy(dst, smaller_dst, smaller_dstlen);
-		*dstlen = smaller_dstlen;
-	}
-	slab_free(NULL, dst2);
 	return (rv);
 }
 

rabin/rabin_dedup.c

@@ -67,7 +67,8 @@
 
 #include "rabin_dedup.h"
 
-#define	FORTY_PCNT(x) (((x)/5 << 1))
+#define	FORTY_PCNT(x) ((x)/5 << 1)
+#define	FIFTY_PCNT(x) ((x) >> 1)
 #define	SIXTY_PCNT(x) (((x) >> 1) + ((x) >> 3))
 
 extern int lzma_init(void **data, int *level, ssize_t chunksize);
@@ -170,11 +171,27 @@ create_dedupe_context(uint64_t chunksize, uint64_t real_chunksize, int rab_blk_s
 
 	ctx->fixed_flag = fixed_flag;
 	ctx->rabin_break_patt = 0;
-	ctx->delta_flag = delta_flag;
 	ctx->rabin_poly_avg_block_size = 1 << (rab_blk_sz + RAB_BLK_MIN_BITS);
 	ctx->rabin_avg_block_mask = ctx->rabin_poly_avg_block_size - 1;
 	ctx->rabin_poly_min_block_size = dedupe_min_blksz(rab_blk_sz);
 	ctx->fp_mask = ctx->rabin_avg_block_mask | ctx->rabin_poly_avg_block_size;
+	ctx->delta_flag = 0;
+
+	/*
+	 * Scale down similarity percentage based on avg block size unless user specified
+	 * argument '-EE' in which case fixed 40% match is used for Delta compression.
+	 */
+	if (delta_flag == DELTA_NORMAL) {
+		if (ctx->rabin_poly_avg_block_size < (1 << 14)) {
+			ctx->delta_flag = 1;
+		} else if (ctx->rabin_poly_avg_block_size < (1 << 16)) {
+			ctx->delta_flag = 2;
+		} else {
+			ctx->delta_flag = 3;
+		}
+	} else if (delta_flag == DELTA_EXTRA) {
+		ctx->delta_flag = 1;
+	}
 
 	if (!fixed_flag)
 		ctx->blknum = chunksize / ctx->rabin_poly_min_block_size;
@@ -356,7 +373,7 @@ dedupe_compress(dedupe_context_t *ctx, uchar_t *buf, ssize_t *size, ssize_t offs
 	j = 0;
 
 	for (i=offset; i<*size; i++) {
-		ssize_t pc[3];
+		ssize_t pc[4];
 		uchar_t cur_byte = buf1[i];
 		uint64_t pushed_out = ctx->current_window_data[ctx->window_pos];
 		ctx->current_window_data[ctx->window_pos] = cur_byte;
@@ -414,7 +431,8 @@ dedupe_compress(dedupe_context_t *ctx, uchar_t *buf, ssize_t *size, ssize_t offs
 			 */
 			if (ctx->delta_flag) {
 				pc[1] = SIXTY_PCNT(j);
-				pc[2] = FORTY_PCNT(j);
+				pc[2] = FIFTY_PCNT(j);
+				pc[3] = FORTY_PCNT(j);
 
 				reset_heap(&heap, pc[ctx->delta_flag]);
 				ksmallest(fplist, j, &heap);
@@ -444,7 +462,8 @@ dedupe_compress(dedupe_context_t *ctx, uchar_t *buf, ssize_t *size, ssize_t offs
 
 			if (j > 1) {
 				pc[1] = SIXTY_PCNT(j);
-				pc[2] = FORTY_PCNT(j);
+				pc[2] = FIFTY_PCNT(j);
+				pc[3] = FORTY_PCNT(j);
 				reset_heap(&heap, pc[ctx->delta_flag]);
 				ksmallest(fplist, j, &heap);
 				cur_sketch =