Flexible I/O Tester development
 help / color / mirror / Atom feed
* [PATCH] Optimize pattern verify
@ 2012-02-07  2:50 Steven Lang
  2012-02-07  8:40 ` Jens Axboe
  0 siblings, 1 reply; 2+ messages in thread
From: Steven Lang @ 2012-02-07  2:50 UTC (permalink / raw)
  To: fio

Similar to the patch last week, this optimizes the pattern verify
operation to use optimized library calls like memcmp(), and only fall
back to byte-by-byte if there is a miscompare to locate it.

This uses the same premise that the pattern is repeated as many times
as possible to do large compares in a single call.  For single byte
pattern, the setup fills the pattern space, and verify assumes it is
full.

Tested by running a script which created an 8k file with 4k bs and,
one byte at a time, tried corrupting the pattern and running a read
pass and verified it still found miscompares across the whole range of
the pattern.  This was done with a pattern length of 1 and 3 bytes.

In performance tests, this was about 8 times more efficient on
verifies than without this patch.

diff --git a/options.c b/options.c
index 8fb93ca..c3fdb56 100644
--- a/options.c
+++ b/options.c
@@ -698,6 +698,14 @@ static int str_verify_pattern_cb(void *data,
const char *input)
 		memcpy(&td->o.verify_pattern[i], &td->o.verify_pattern[0], i);
 		i *= 2;
 	}
+	if (i == 1) {
+		/*
+		 * The code in verify_io_u_pattern assumes a single byte pattern
+		 * fills the whole verify pattern buffer.
+		 */
+		memset(td->o.verify_pattern, td->o.verify_pattern[0],
+		       MAX_PATTERN_SIZE);
+	}

 	td->o.verify_pattern_bytes = i;

diff --git a/verify.c b/verify.c
index 9ee3bc4..eb8eddc 100644
--- a/verify.c
+++ b/verify.c
@@ -322,14 +322,27 @@ static int verify_io_u_pattern(struct
verify_header *hdr, struct vcont *vc)
 	struct io_u *io_u = vc->io_u;
 	char *buf, *pattern;
 	unsigned int header_size = __hdr_size(td->o.verify);
-	unsigned int len, mod, i;
+	unsigned int len, mod, i, size, pattern_size;

 	pattern = td->o.verify_pattern;
+	pattern_size = td->o.verify_pattern_bytes;
+	if (pattern_size <= 1)
+		pattern_size = MAX_PATTERN_SIZE;
 	buf = (void *) hdr + header_size;
 	len = get_hdr_inc(td, io_u) - header_size;
-	mod = header_size % td->o.verify_pattern_bytes;
+	mod = header_size % pattern_size;
+
+	for (i = 0; i < len; i += size) {
+		size = pattern_size - mod;
+		if (size > (len - i))
+			size = len - i;
+		if (memcmp(buf + i, pattern + mod, size))
+			// Let the slow compare find the first mismatch byte.
+			break;
+		mod = 0;
+	}

-	for (i = 0; i < len; i++) {
+	for (; i < len; i++) {
 		if (buf[i] != pattern[mod]) {
 			unsigned int bits;

--

^ permalink raw reply related	[flat|nested] 2+ messages in thread

* Re: [PATCH] Optimize pattern verify
  2012-02-07  2:50 [PATCH] Optimize pattern verify Steven Lang
@ 2012-02-07  8:40 ` Jens Axboe
  0 siblings, 0 replies; 2+ messages in thread
From: Jens Axboe @ 2012-02-07  8:40 UTC (permalink / raw)
  To: Steven Lang; +Cc: fio

On 02/07/2012 03:50 AM, Steven Lang wrote:
> Similar to the patch last week, this optimizes the pattern verify
> operation to use optimized library calls like memcmp(), and only fall
> back to byte-by-byte if there is a miscompare to locate it.
> 
> This uses the same premise that the pattern is repeated as many times
> as possible to do large compares in a single call.  For single byte
> pattern, the setup fills the pattern space, and verify assumes it is
> full.
> 
> Tested by running a script which created an 8k file with 4k bs and,
> one byte at a time, tried corrupting the pattern and running a read
> pass and verified it still found miscompares across the whole range of
> the pattern.  This was done with a pattern length of 1 and 3 bytes.
> 
> In performance tests, this was about 8 times more efficient on
> verifies than without this patch.

Great, applied! Thanks.

-- 
Jens Axboe


^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2012-02-07  8:41 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2012-02-07  2:50 [PATCH] Optimize pattern verify Steven Lang
2012-02-07  8:40 ` Jens Axboe

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox