Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date: Wed, 13 Jun 2012 17:12:44 +0400
From: Solar Designer <solar@...nwall.com>
To: john-dev@...ts.openwall.com
Subject: Re: Lukas - status report #8

Lukas -

On Wed, Jun 13, 2012 at 03:25:28AM +0200, Lukas Odzioba wrote:
> THIS WEEK:
> I made following formats (based on pbkdf2-sha512):
> -grub-cuda:
> 1900 c/s on GTX570, 1250 c/s on GTX460
> -grub - openssl based 41c/s on bulls 1 thread
> 
> all tests on default 12k rounds - thats 24k sha512 calls

So that's 45 million of SHA-512's per second, whereas we know that the
GTX 570 can do about 70 million.  This is fine, yet perhaps there's
significant room for optimization.

Was there any demand for cracking GRUB's hashes, though?  I haven't
heard of any.  I understand that the OpenCL code for PBKDF2-SHA-512 will
be reusable, though.

> NEXT WEEK:
> Priority order:
> -finish grub-opencl
> -fix grub OpenMP issue
> -investigate and fix problem with TS reported by magnum
> -work  OSX 10.8 format
> -try grub cpu version without openssl - it should be faster

I don't mind, but here's a concern: you appear to be avoiding the more
difficult portions of work, whereas I had expected that those would be
the primary focus of your GSoC work this summer - that is, a difficulty
increase compared to your GSoC 2011 project.  Here are some things I
think you can/should work on this summer:

Working md5crypt on HD 7970 (I think yours still fails there, doesn't it?)

Faster md5crypt.  Yours delivers about the same c/s rate as phpass at
$P$9, whereas it should be about twice faster (since it does twice fewer
MD5 computations).  (Both CUDA and OpenCL need improvement.)

Faster phpass in OpenCL.  (I think yours still under-performs by a factor
of 2 on HD 7970  Your CUDA code for phpass is already near-optimal.)

Bitslice DES on GPU, both OpenCL and CUDA.  DES-based crypt(3) is
semi-fast, so it's inbetween your and myrice's GSoC projects in this
sense, but BSDI-style one is slow (under typical settings).  And indeed
the DES code itself should be reusable for other DES-based formats,
including fast ones.

I listed the above roughly in priority order.  Bitslice DES is listed
last because it might also be picked up and worked on by myrice.

Thanks,

Alexander

Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.