(usagi-users 03112) Re: ipv6_addr_cmp performance patch

[Brian Haley <Brian.Haley@xxxxxx>]

Craig, Dave wrote:

> Please consider the attached patch for incorporation into the next Linux
> 2.6 release.  I found that using the __builtin_memcmp GCC intrinsic for
> doing IPv6 address comparison a lot more slow than integer comparison.
> In significant SCTP workloads I found address comparison taking roughly
> 10% of the CPU time.  After changing the address comparison to an
> integer-based approach the comparison consumed 5% of the CPU time.

This is what IN6_ARE_ADDR_EQUAL() does in netinet/in.h in user-space.

> I also microbenchmarked the comparison on my architecture, a Pentium 4,
> and found the integer comparison was over 3x faster than the string
> comparison.  The microbenchmark consisted of a series of address
> comparisons where 20% matched and 80% that didn't match.  I looked at
> addresses that differed in the first, last, and random byte.

How does performance compare on a 64-bit architecture?  If you send me a 
pointer I can at least try IA64, maybe Opteron.

Maybe a more general question is why none of the macros/inlines have 
optimized 64-bit versions, which would make this 2 integer compares. 
There must be a reason the below isn't defined?

struct in6_addr
{
        union
        {
                __u8            u6_addr8[16];
                __u16           u6_addr16[8];
                __u32           u6_addr32[4];
+		__u64		u6_addr64[2];
        } in6_u;
#define s6_addr                 in6_u.u6_addr8
#define s6_addr16               in6_u.u6_addr16
#define s6_addr32               in6_u.u6_addr32
+#defind s6_addr64		in6_u.u6_addr64
};

-Brian