Date: Fri, 18 Apr 2014 00:56:06 +0400
From: Solar Designer <solar@...nwall.com>
To: discussions@...sword-hashing.net
Subject: Re: [PHC] Lyra2 initial review

On Thu, Apr 17, 2014 at 04:36:55PM -0400, Bill Cox wrote:
> On Thu, Apr 17, 2014 at 4:10 PM, Solar Designer <solar@...nwall.com> wrote:
> > This assumes a total of 7 reads plus writes per location, which I'm not
> > sure is correct.  Looking at the algorithm, it might be even more, but
> > that would push the memory bandwidth usage even higher, which would be
> > surprisingly high.  It might be 4 r/w per iteration in Filling Loop, and
> > 5 r/w per iteration in Visitation Loop, so 9 total.  If this is correct,
> > we'd see 13.68 GiB/s here.  This might be realistic because, if I
> > understand correctly, the same rand is reused with and without rotW(),
> > which halves the amount of crypto-like processing per memory access even
> > compared to reduced round TwoCats and yescrypt.  (If one round of
> > pwxform is crypto-like, whereas one bit-rotate is not.)
> 
> I think for DRAM bandwidth, 7 is the right number.

I think you're correct.  I mistakenly counted the reads from prev,
whereas we are not counting the equivalent of those for TwoCats and
yescrypt DRAM bandwidth usage.

> > BTW, yescrypt at zero rounds of pwxform might well be a reasonable thing
> > to use, if we're to maximize memory bandwidth usage while not even
> > trying to keep any compute time hardening (besides the occasional Salsa20).
> > I just haven't studied it enough to recommend it as an option, and I am
> > not aware of a use case where zero compute time hardening would be good.
> > When making the rounds count tunable (and encoding it in the flags
> > field), I am not sure if I should make zero a supported value.
> 
> I think having a mode that is not explicitly compute-time hardened would be
> a good thing.  My SkinnyCat subset of TwoCats has no multiplication chains,
> and only has memory bandwidth to force attackers to take a while.  For
> machines without a dedicated multiplier, this might be a better mode for
> them to run.

If Salsa20 (almost) does not count in terms of ASIC compute time
hardening, then we sort of have that mode in yescrypt already, by simply
not setting the YESCRYPT_PWXFORM flag.  However, then Salsa20 rounds
count would need to be made tunable, if we want to maximize memory
bandwidth usage in that mode.

For CPUs without fast enough multipliers, I think (tweaked) yescrypt
should preferably retain bcrypt-like GPU unfriendliness.  I am thinking
to treat S_SIMD*S_P = 0 as requesting this special mode: it's kind of
less than one normal pwxform 64-bit lane, so I could use a single 32-bit
lane without the multiplies then, and would potentially have ~2x to ~4x
better GPU resistance than bcrypt on those older CPUs (by using only 2
instead of 4 S-boxes, and/or by making their combined size 8 KiB instead
of 4 KiB).  I am assuming that CPUs without fast multipliers typically
also don't need more than 32 bits of parallelism, and that any 64-bit
CPU would have fast multipliers (at least one).

> I don't know what skipping pwxform calls does to security, though.  I'm
> sure you can deal with that.

I think this has no undesired side-effects, other than the expected
removal of GPU resistance and compute-time hardening, but it does feel
like we'd be getting (too?) close to the edge of the cliff.  On the
other hand, it'd be nice to have yescrypt reviewed and attacked in this
cut-down form (as well as in full form) during PHC.

Alexander

Powered by blists - more mailing lists