Re: HalfSipHash Acceptable Usage

[Eric Dumazet]

On Wed, 2016-12-21 at 15:42 +0100, Jason A. Donenfeld wrote:
> Hi Eric,
> 
> I computed performance numbers for both 32-bit and 64-bit using the
> actual functions in which talking about replacing MD5 with SipHash.
> The basic harness is here [1] if you're curious. SipHash was a pretty
> clear winner for both cases.
> 
> x86_64:
> [    1.714302] secure_tcpv6_sequence_number_md5# cycles: 102373398
> [    1.747685] secure_tcp_sequence_number_md5# cycles: 92042258
> [    1.773522] secure_tcpv6_sequence_number_siphash# cycles: 70786533
> [    1.798701] secure_tcp_sequence_number_siphash# cycles: 68941043
> 
> x86:
> [    1.635749] secure_tcpv6_sequence_number_md5# cycles: 106016335
> [    1.670259] secure_tcp_sequence_number_md5# cycles: 95670512
> [    1.708387] secure_tcpv6_sequence_number_siphash# cycles: 105988635
> [    1.740264] secure_tcp_sequence_number_siphash# cycles: 88225395
> 
> >>> 102373398 > 70786533
> True
> >>> 92042258 > 68941043
> True
> >>> 106016335 > 105988635
> True
> >>> 95670512 > 88225395
> True
> 
> While MD5 is probably faster for some kind of large-data
> cycles-per-byte, due to its 64-byte internal state, SipHash -- the
> "Sip" part standing "Short Input PRF" -- is fast for shorter inputs.
> In practice with the functions we're talking about replacing, there's
> no need to hash 64-bytes. So, SipHash comes out faster and more
> secure.
> 
> I also haven't begun to look focusedly at the assembly my SipHash
> implemention is generating, which means there's still window for even
> more performance improvements.
> 
> Jason
> 
> 
> [1] https://git.zx2c4.com/linux-dev/tree/net/core/secure_seq.c?h=siphash-bench#n194

Now I am quite confused.

George said :

> Cycles per byte on 1024 bytes of data:
>                       Pentium Core 2  Ivy
>                       4       Duo     Bridge
> SipHash-2-4           38.9     8.3     5.8
> HalfSipHash-2-4               12.7     4.5     3.2
> MD5                    8.3     5.7     4.7


That really was for 1024 bytes blocks, so pretty much useless for our
discussion ?

Reading your numbers last week, I thought SipHash was faster, but George
numbers are giving the opposite impression.

I do not have a P4 to make tests, so I only can trust you or George.

Thanks.