phc-discussions - Re: [PHC] BlaMka loses entropy

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: Mon, 19 Oct 2015 22:02:03 +0300
From: Solar Designer <solar@...nwall.com>
To: discussions@...sword-hashing.net
Subject: Re: [PHC] BlaMka loses entropy

On Mon, Oct 19, 2015 at 11:17:44AM -0700, Bill Cox wrote:
> - Poor in-cache performance (about 3X slower than TwoCats)

Are you referring to your ~1ms running time measurements at 4 MiB?

If so, that's a non-issue.  I am puzzled why you keep bringing it up.

I have yet to see any benchmarks showing Argon2 run much slower than
TwoCats in terms of password hashing throughput (hashes per second on a
fully loaded CPU).  Do you have such benchmark results maybe?

> Using MAXFORM on the scalar unit would solve 4 of my top 5 concerns.

I'd say 3 out of 3:

> - Too much parallelism
> - Poor compute-time hardening (due to parallelism)
> - Somewhat poor GPU resistance.

These are real issues.

> The
> poor in-cache performance is not easily fixed, since Argon2's state busts
> out of the SIMD unit registers and lives in L1 cache.  For low-memory
> in-cache memory hashing, I plan to replace the Argon2 block hash with most
> likely either the one from TwoCats or preferably Yescrypt if it is fast
> enough.

FWIW, the final yescrypt will write at least as much data to L1 cache as
Argon2 does (when running with PWXrounds=3; or 4 times that much data
when running with the default PWXrounds=6).  A notable difference is
that those writes won't be temporary for the current block's processing,
but will also affect further blocks.  Per my recent testing, sequential
writes to L1 cache are essentially free on modern x86 CPUs, even on
Bulldozer (where they go to WCC and L2), so in a sense it's wasteful not
to do those.  (OTOH, Argon2's are wasteful in that Argon2 does not take
full advantage of them.  The ds flavor could possibly use them to update
the S-boxes, much like yescrypt will.)

If you'd like to run those password hashing throughput benchmarks I
suggested above, I'd be interested if you include yescrypt-0.8+ in those
as well (to be made public soon, or I may share it with you now), at the
two PWXrounds settings I mentioned (3 and 6).

Alexander