SipHash benchmarks: I will use HalfSipHash31

[Sam Trenholme <lua@...>]

I have run a number of benchmarks with Lunacy, my fork of Lua 5.1, to 
see how much changing the SipHash variant used affects performance, for 
both 32-bit (386) and 64-bit (x86_64) binaries.

Conclusion: I will use HalfSipHash31 as the hash compression algorithm, 
for both 32-bit and 64-bit builds of Lunacy.

The binaries have been compiled using GCC 8.3.1, in CentOS 8, using an 
older Core Duo T8100 chip from 2008.  The benchmark consisted of loading 
and processing a bunch of COVID-19 data in to large tables taking up 550 
(32-bit) or 750 megs (64-bit) of memory.  This real-world benchmark (it 
is the exact same code I use to build an entire COVID-19 tracking 
website) was done multiple times, to minimize speed fluctuations from 
outside factors, against the following setups:

• “Lunacy32”, which is a 32-bit compile of Lua 5.1
• “Lunacy64”, a 64-bit compile of same

And the following string hash compression functions:

• “noSipHash”: Lua’s default hash compressor
• “SipHash42”: 64-bit SipHash with 2 rounds during input processing, 
followed by 4 rounds after input ends.
• “SipHash31”: 64-bit SipHash with 1 round during input processing, 
followed by 3 rounds after input ends.
• “SipHalf31”: 32-bit HalfSipHash with 1 round during input processing, 
followed by 3 rounds after input ends.

Here are the results, where lower numbers are better (less time needed 
to run the benchmark):

lunacy64-noSipHash 197.801
lunacy64-sipHash31 203.457
lunacy64-SipHalf31 203.507
lunacy64-sipHash42 210.043
lunacy32-noSipHash 240.898
lunacy32-SipHalf31 246.995
lunacy32-sipHash31 265.916
lunacy32-sipHash42 270.226

HalfSipHash31 is as fast as full SipHash31 on 64-bit CPUs, while being 
quite a bit faster for 32-bit binaries compared to 64-bit sipHash.

HalfSipHash31 is only 2.5% slower on 32-bit machines (compared to Lua’s  
“stock” hash); it is only 2.9% slower on 64-bit machines.

Since HalfSiphash only has a 64-bit key, it is not fully secure in cases 
where an attacker has significant computing resources to brute force the 
key, and where the attacker is able to see the hash compression values 
for strings.  However, in the case of using HalfSipHash31 as a hash 
compression string, an attacker will not be able to see the hash 
compression values; to do so would require viewing the memory used by a 
Lua process -- if an attacker has that level of access, there are other 
far more devistating attacks they can perform.

-- Sam