[openssl-dev] use SIPhash for OPENSSL_LH_strhash?

[Richard Levitte]

In message <1e19cdfea8224717b3eee11e2d8acda5 at usma1ex-dag1mb1.msg.corp.akamai.com> on Wed, 11 Jan 2017 03:13:39 +0000, "Salz, Rich" <rsalz at akamai.com> said:

rsalz> The needs for OpenSSL's LHASH are exactly what SipHash was designed for: fast on short strings.
rsalz> OpenSSL's hash currently *does not* call MD5 or SHA1; the MD5 code is commented out.
rsalz> Yes, performance tests would greatly inform the decision.

Done, using the reference siphash implementation.

https://github.com/levitte/openssl/tree/test-string-hashes

A run on my laptop gave these results:

    : ; ./util/shlib_wrap.sh apps/openssl speed siphash lhash
    Doing lhash for 3s on 16 size blocks: 27635188 lhash's in 3.00s
    Doing lhash for 3s on 64 size blocks: 6934726 lhash's in 3.00s
    Doing lhash for 3s on 256 size blocks: 1698489 lhash's in 3.00s
    Doing lhash for 3s on 1024 size blocks: 431185 lhash's in 3.00s
    Doing lhash for 3s on 8192 size blocks: 53868 lhash's in 3.00s
    Doing lhash for 3s on 16384 size blocks: 27041 lhash's in 3.00s
    Doing siphash for 3s on 16 size blocks: 22488748 siphash's in 3.00s
    Doing siphash for 3s on 64 size blocks: 10485674 siphash's in 3.00s
    Doing siphash for 3s on 256 size blocks: 3320898 siphash's in 3.00s
    Doing siphash for 3s on 1024 size blocks: 894647 siphash's in 3.00s
    Doing siphash for 3s on 8192 size blocks: 114170 siphash's in 3.00s
    Doing siphash for 3s on 16384 size blocks: 57151 siphash's in 3.00s
    OpenSSL 1.1.1-dev  xx XXX xxxx
    built on: reproducible build, date unspecified
    options:bn(64,64) rc4(16x,int) des(int) aes(partial) idea(int) blowfish(ptr) 
    compiler: gcc -DDSO_DLFCN -DHAVE_DLFCN_H -DOPENSSL_THREADS -DOPENSSL_NO_STATIC_ENGINE -DOPENSSL_PIC -DOPENSSL_IA32_SSE2 -DOPENSSL_BN_ASM_MONT -DOPENSSL_BN_ASM_MONT5 -DOPENSSL_BN_ASM_GF2m -DSHA1_ASM -DSHA256_ASM -DSHA512_ASM -DRC4_ASM -DMD5_ASM -DAES_ASM -DVPAES_ASM -DBSAES_ASM -DGHASH_ASM -DECP_NISTZ256_ASM -DPADLOCK_ASM -DPOLY1305_ASM -DOPENSSLDIR="\"/usr/local/ssl\"" -DENGINESDIR="\"/usr/local/lib/engines-1.1\""  -DDEBUG_UNUSED -Wswitch -DPEDANTIC -pedantic -Wno-long-long -Wall -Wsign-compare -Wmissing-prototypes -Wshadow -Wformat -Wtype-limits -Werror -Wa,--noexecstack
    The 'numbers' are in 1000s of bytes per second processed.
    type             16 bytes     64 bytes    256 bytes   1024 bytes   8192 bytes  16384 bytes
    lhash           147387.67k   147940.82k   144937.73k   147177.81k   147095.55k   147679.91k
    siphash         119939.99k   223694.38k   283383.30k   305372.84k   311760.21k   312120.66k

So it seems that for short strings, OPENSSL_LH_strhash (*) wins some,
while siphash wins big for larger strings.

I have no idea how they compare with regard to distribution (which,
considering I ask for the same size output from both, should be the
main factor that affects the sensitivity to hash flooding)...

Our use of OPENSSL_LH_strhash() is with configuration sections and
names, ASN.1 object names and the function names in the openssl app.
All our other uses of lhash use their own hashing functions, and are
usually very short still (definitely less than 16 bytes).

My conclusion is that performance-wise, siphash doesn't give us any
advantage over OpenSSL_LH_strhash for our uses.

Cheers,
Richard

(*) Strictly speaking, it's a modified version that takes a length and
tolerates all 8-bit bytes, including 0x00.

-- 
Richard Levitte         levitte at openssl.org
OpenSSL Project         http://www.openssl.org/~levitte/