phc-discussions - Re: [PHC] Updated tests document

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, 16 Apr 2015 17:05:30 +0300
From: Solar Designer <solar@...nwall.com>
To: discussions@...sword-hashing.net
Subject: Re: [PHC] Updated tests document

Hi Milan,

On Thu, Apr 16, 2015 at 08:53:00AM +0200, Milan Broz wrote:
> I tried to hack something, it is not perfect but it could provide some overview,
> but please take it with a grain of (non-crypto) salt :-)

This is very helpful.  Thank you!

You're benchmarking via the PHS() interface, right?  For authentication
servers, a mode with reuse of a previous memory allocation also makes
sense.  We should keep in mind that schemes demonstrating high hashes
per second rates may be made even faster on such server setups, as well
as during attack.  Improvements e.g. from 2000 to 3000 hashes/second at
1 MB are possible in this way.

yescrypt offers an API for this.  I think others currently don't.  It's
probably fair not to add this sort of benchmarks yet (and it would be
too much work for you to create the necessary APIs or hacks), but it's
something to keep in mind.

> - using memory table from Steve (128k/1M/16M/128M)

OK.

> - test forks N processes, parent waits 60 seconds (in this tests) and then
> *waits* for all children to finish the last hash (this time is measured).
> I am using SHM quite stupid way here to sync it, but it works it this case.

Sounds fine, but I didn't look at the code.

> (Pufferfish and Argon1 on 100MiB run time exceeds 60 seconds, so measurement
> is really not reliable here.)

Ouch.  Overall, per these results I think Pufferfish and Argon1 (without
AES-NI) are out as potential winners for this use case.  They might
still be used like that when a higher hashes/second rate is not needed
anyway, but I think we must not select them as PHC winners as they are
not general enough.  (And of course Pufferfish needs a bug fix before it
can be used above 2 MB.)

battcrypt and Catena-Butterfly are borderline.  I thought battcrypt
would perform better.  Maybe there's room for code optimization.

> (Moreover, all hashes all the same but this should not be a problem.)

Do you mean you're repeatedly hashing the same passwords?  This may
potentially affect the speed figures.  I suggest that you correct this.

For example, yescrypt's userom.c benchmark uses a time-dependent
sequence of integers as passwords.

> - it increases # processes up to 8, > 4 cpu it should not add anything.

Actually, > 4 CPU may increase speeds further if one instance of the
hash doesn't have enough instruction-level parallelism in it and going
for 2 concurrent threads/core would not exceed a cache size.  OTOH, it
may also decrease speeds if we're just crossing a cache size there.
So I appreciate you including such benchmarks as well.  It's useful to
know that all of the benchmarked schemes behave reasonably in the 4 to 8
threads range.

> I also added yescrypt with reduced 2 PWXrounds to repo, but I guess it makes sense only
> for "normalized" test chart...

I think it's of as much relevance to this one as well.

Also, I'd like to see Makwa added, perhaps to the 128 KB chart.

Thanks again,

Alexander