phc-discussions - Re: [PHC] BSTY - yescrypt-based cryptocoin

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date: Thu, 11 Sep 2014 11:24:05 +0400
From: Solar Designer <solar@...nwall.com>
To: discussions@...sword-hashing.net
Subject: Re: [PHC] BSTY - yescrypt-based cryptocoin

On Wed, Sep 10, 2014 at 07:15:43AM +0400, Solar Designer wrote:
> On Tue, Sep 09, 2014 at 03:45:28PM -0400, Bill Cox wrote:
> > I did a quick comparison of TwoCats and Yescrypt when doing 2MiB
> > hashes.  Yescrypt maxes out my machine at about 3,100 hashes per
> > second using 8 threads, which gives the best performance.  TwoCats
> > maxes out at about 3,800 similar sized hashes on 3 threads with 2
> > multiplications per inner loop, which gives the best performance.
> > However, Yescrypt is doing something like 2.3 memory read/writes per
> > location vs TwoCat's 2.  The difference is basically in the noise.
> 
> Yeah.  To add one more number, yescrypt with pwxform rounds count
> reduced from 6 to 2 does 4400 hashes per second on the same i7-4770K
> where it does 3400 hashes per second with the default of 6.  I guess
> this might correspond to more than 3800 on your machine. ;-)  But I
> think the default usually works better, since it provides much more
> compute latency hardening:
> 
> 6*3400/(2*4400) = 2.3 times more hardening

I realized I should have tried other thread counts when benchmarking
yescrypt with reduced pwxform rounds count, because at only 2 MiB per
hash L3 cache plays a significant role.  It turns out that although with
the default of 6 pwxform rounds 8 threads were a bit faster than
4 threads (3400 vs. 3150 hashes per second), with only 2 pwxform rounds
4 threads win, with between 6300 and 6700 hashes per second (the speed
becomes somewhat unstable, perhaps because of the structure of L3 cache
and the effect of context switches).  So the reduction in compute
latency hardening is much smaller than I previously thought:

6*3400/(2*6500) = 1.57

This actually might be a better fit for such use, because it uses much
more bandwidth (to cache or RAM or whatever memory the attacker provides):

6500/3400 = 1.91

So perhaps I should in fact export the pwxform rounds count as a
parameter, perhaps with some granularity.

Also interesting is what happens with 1 MiB per hash:

6 rounds, 4 threads: 6300
6 rounds, 8 threads: 7300
2 rounds, 4 threads: 13400
2 rounds, 8 threads: 14550 to 15400 (unstable speed)

As expected, this makes 8 threads the optimal choice again.  The speed
and bandwidth usage difference between 6 and 2 rounds improves to 2x+.
Of course, like before this is countered by the 3x reduction in compute
hardening per hash computed (so up to 1.5x overall).

When AVX2 code is written, it will improve speeds at 6 rounds slightly
(but less so at 2 rounds).

Alexander