JFFS3 & performance

[Joakim Tjernlund]

> On Thu, 16 December 2004 20:15:00 +0100, Jörn Engel wrote:
> >
> > PS: Now you've done it.  I'll implement crc24 and crc16 and benchmark
> > them against adler32.  Darn you!
>
> Testcase going through 45MB of data in chunks of 4k.  Machine is
> PIII-1166 with warm caches.
>
> Doing three runs and discarding the fastest and slowest ones:
>
> crc32:
> real    0m0.214s
> user    0m0.133s
> sys     0m0.076s
>
> adler32:
> real    0m0.128s
> user    0m0.061s
> sys     0m0.066s
>
> crc24:
> real    0m0.969s
> user    0m0.882s
> sys     0m0.073s
>
> crc16:
> real    0m0.382s
> user    0m0.312s
> sys     0m0.061s
>
>
> Looks like those cold hard numbers beat the crap out of my previous
> argument.  So unless someone can seriously optimize below functions,
> just pick adler32.

A table driven crc24/crc16 is faster, I think.

>
>
> uint32_t crc24(uint32_t crc, const void *_s, size_t len)
> {
> 	const char *s = _s;
> 	uint32_t ret = crc;
> 	strlen(s);

Why strlen()?

> 	for (; len; len--,s++) {
> 		ret <<= 8;
> 		ret += *s;
> 		ret %= 0xfffffd;
> 	}
> 	return ret;
> }

This is probably faster(on PPC).
Depens on gcc version as well.

uint32_t crc24(uint32_t crc, const void *_s, size_t len)
{
	const char *s = _s-1;
	uint32_t ret = crc;
	if (len)
		do {
			ret <<= 8;
			ret += *++s;
			ret %= 0xfffffd;
		} while (--len);
	return ret;
}
>
> uint32_t crc16(uint32_t crc, const void *_s, size_t len)
> {
> 	const uint16_t *s = _s;
> 	uint32_t ret = crc;
> 	for (; len>1; len-=2,s++) {
> 		ret <<= 16;
> 		ret += *s;
> 		ret %= 65521;
> 	}
> 	return ret;
> }
>
>
> Just in order to be complete, here is a variant of reiserfs' r5 hash.
> Noone should seriously use it for error detection, but the results are
> nice for comparison:
> real    0m0.207s
> user    0m0.134s
> sys     0m0.067s
>
>
> Jörn
>
> --
> "Translations are and will always be problematic. They inflict violence
> upon two languages." (translation from German)