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  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: Sat, 14 Feb 2015 01:48:32 +0300
From: Solar Designer <solar@...nwall.com>
To: discussions@...sword-hashing.net
Subject: Re: [PHC] Tradeoff cryptanalysis of Catena, Lyra2, and generic memory-hard functions

On Fri, Feb 13, 2015 at 07:43:50PM -0200, Marcos Simplicio wrote:
> Well, maybe we did something wrong in our benchmarks then, because our
> implementations are comparisons done in v2 are faster than Yescrypt with
> minimal parameters for both functions, both with 1 thread and with
> multiple threads...

You have some impressive benchmarks in the current v3 submission!
It's very helpful that you included multiple PHC finalists there.

It may well be the case that Lyra2 is now faster than yescrypt, at
yescrypt's current default PWXrounds = 6.  (yescrypt's use of
PWXrounds = 6 may be justified in terms of ASIC area-time cost, though -
in the special case that an ASIC would be compute latency rather than
memory latency bound, which with yescrypt's use of integer
multiplication may sometimes be the case.)

Naturally, I would be interesting in adding yescrypt benchmarks for
lower PWXrounds as well, down to 1.

Also, unless I missed it, you don't appear to have ever benchmarked
these at the maximum hardware thread count supported by your CPUs.
You mention "12 cores" (is this 12 hardware threads? or 12 cores across
two CPU chips, and 24 hardware threads total?), but somehow you only
included benchmarks for up to p = 4?

yescrypt is tuned to maximize resource usage (balanced between overall
CPU instruction issue rate across all hardware threads, and total
memory bandwidth) when running the maximum number of threads that are
supported in hardware (this matches the case of an authentication server
bumping into its request rate capacity during a spike).  It tries not to
bump into memory bandwidth prematurely, which would mean that the CPU
cores are not fully used to maximize latency in attack implementations.

> However, these benchmarks refer to yescrypt v0. Are your comments
> referring to v1?

I'm not Bill, but I guess Bill was referring to older Lyra2 being
unaware that you speed it up.

As to yescrypt, its performance has not changed much.  Only the
pre-hashing and S-box initialization have changed in performance, and
these usually correspond to a very small part of the total run time.

> I agree that going below 1 round in Lyra2 is not a great option, but I'm
> not sure about the mixing ability of yescrypt's dedicated function when
> compared with Blake2. I mean, for a internal of 1024 bits, 1 round of
> Blake2 ensures that every bit of the internal and external states depend
> on every input bit. Does yescrypt's function do the same? (note: this is
> really a question, not a "provocation" of any sort). Because if it does
> not, then in theory Blake2 could be reduced even more to match
> yescrypt's diffusion capabilities (although personally I would not
> recommend it).

yescrypt's BlockMix_pwxform has no diffusion between pwxform's 64-bit
lanes until it reaches the final sub-block, where Salsa20/8 is used to
mix the lanes.

I don't know Lyra2 specifics to the same extent as I know yescrypt's,
but it is quite possible that you could go below 1 round of BLAKE2 for
all but the last sub-block as well.  Of course, you'd need to review
this very carefully to ensure that sufficient data dependencies remain
(to avoid attack shortcuts), etc.  I doubt it's a good idea to do this
now, though.  (And it's too late for PHC.)  A Lyra2 successor like that
would probably lose a lot of people's confidence in it if you did that.

In yescrypt, the lanes are not mixed until sub-block end not so much to
speed things up (this is lost by setting PWXrounds > 1 anyway, as is now
the default), but primarily to make yescrypt extremely scalable in
terms of instruction-level, SIMD, and gather loads parallelism.  This is
to allow for efficient implementations on architectures very different
from current CPUs.  If in a Lyra2 successor your rationale would only be
speed, it's probably not enough of a reason to go for that.

Alexander

Powered by blists - more mailing lists