Re: [PATCH] dcache: better name hash function

[Eric Dumazet <eric.dumazet@gmail.com>]

Eric Dumazet a écrit :
> unsigned int fold2(unsigned hash)
> {
> 	return ((unsigned long long)hash * 211) >> 32;
> }
> 

I tried this reciprocal thing with 511 and 1023 values and got on a PIII 550 MHz, gcc-3.3.2 :

# ./hashtest 100000 511 
jhash_string         0.033123       1.01       234   1.06
fnv32                0.033911       1.02       254   1.38
# ./hashtest 1000000 511
jhash_string         0.331155       1.00      2109   1.10
fnv32                0.359346       1.00      2151   1.65
# ./hashtest 10000000 511
jhash_string         3.383340       1.00     19985   1.03
fnv32                3.849359       1.00     20198   1.53

# ./hashtest 100000 1023
jhash_string         0.033123       1.03       134   1.01
fnv32                0.034260       1.03       142   1.32
# ./hashtest 1000000 1023
jhash_string         0.332329       1.00      1075   1.06
fnv32                0.422035       1.00      1121   1.59
# ./hashtest 10000000 1023
jhash_string         3.417559       1.00     10107   1.01
fnv32                3.747563       1.00     10223   1.35


511 value on 64bit, and 1023 on 32bit arches are nice because
hashsz * sizeof(pointer) <= 4096, wasting space for one pointer only.

Conclusion : jhash and 511/1023 hashsize for netdevices,
no divides, only one multiply for the fold.