From 3d8f3ada1c91fb77017f6b9396df8358ff54baeb Mon Sep 17 00:00:00 2001
From: Moinak Ghosh <moinakg@gmail.com>
Date: Wed, 30 Jan 2013 22:41:13 +0530
Subject: [PATCH] Improve Deduplication performance by another 95%. Start
 sliding window scanning near minimum chunk size boundaries to avoid scanning
 whole chunk.

---
 rabin/rabin_dedup.c | 8 ++++++++
 rabin/rabin_dedup.h | 4 ++++
 2 files changed, 12 insertions(+)

diff --git a/rabin/rabin_dedup.c b/rabin/rabin_dedup.c
index 16a0e51..6492dbf 100755
--- a/rabin/rabin_dedup.c
+++ b/rabin/rabin_dedup.c
@@ -416,6 +416,12 @@ dedupe_compress(dedupe_context_t *ctx, uchar_t *buf, uint64_t *size, uint64_t of
 		return (0);
 	}
 
+	/*
+	 * Start our sliding window at a fixed number of bytes before the min window size.
+	 * It is pointless to slide the window over the whole length of the chunk.
+	 */
+	offset = ctx->rabin_poly_min_block_size - RAB_WINDOW_SLIDE_OFFSET;
+	length = offset;
 	for (i=offset; i<j; i++) {
 		uint64_t pc[4];
 		uint32_t cur_byte = buf1[i];
@@ -498,6 +504,8 @@ dedupe_compress(dedupe_context_t *ctx, uchar_t *buf, uint64_t *size, uint64_t of
 			last_offset = i+1;
 			length = 0;
 			if (*size - last_offset <= ctx->rabin_poly_min_block_size) break;
+			length = ctx->rabin_poly_min_block_size - RAB_WINDOW_SLIDE_OFFSET;
+			i = i + length;
 		}
 	}
 
diff --git a/rabin/rabin_dedup.h b/rabin/rabin_dedup.h
index 2aeb4a3..e6d4bbb 100644
--- a/rabin/rabin_dedup.h
+++ b/rabin/rabin_dedup.h
@@ -76,6 +76,10 @@
 #define	RAB_BLK_MASK (((1 << RAB_BLK_MIN_BITS) - 1) >> 1)
 #define	RAB_BLK_AVG_SZ(x) (1 << ((x) + RAB_BLK_MIN_BITS))
 
+// The sliding window starts at min window size - this offset. It is needless
+// to slide the window over every byte in the chunk.
+#define	RAB_WINDOW_SLIDE_OFFSET	(256)
+
 // Minimum practical chunk size when doing dedup
 #define	RAB_MIN_CHUNK_SIZE (1048576L)