Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date: Fri, 25 May 2012 00:25:53 +0200
From: magnum <>
Subject: Re: Can Excessive Rounds make Password cracking Infeasable

On 05/24/2012 11:03 PM, Solar Designer wrote:
> On Thu, May 24, 2012 at 10:49:41PM +0200, magnum wrote:
>> FWIW, some very quick tests on Solar's test gear with the hash above:
>> CPU (8 cores): 4.75 c/s
>> GTX580: 4300 c/s
>> HD7970: 6000 c/s
> These GPU speeds are not realistic.  There must be a bug or just lack of
> support for non-default numbers of iterations for this hash type (this
> is something for us to correct if so!)

That support is there. It turns out I did the CPU test above on just one
core but the figures are still not plausible.

I now modified the formats with an actual test vector with 391939
rounds, so we see that they are doing it right. This also give me
correct c/s figures directly (I had to recalculate the ones above as
discussed on john-dev). For some reason, CUDA now took forever so I
modified the self-tests to only do one test instead of a subset of
max_keys_per_crypt (I hardcoded max=1). New figures:

Benchmarking: crypt SHA-512 (rounds=391939) [OpenSSL 64/64]... (8xOMP) DONE
Raw:    26.3 c/s real, 3.3 c/s virtual

Benchmarking: cryptsha512-cuda (rounds=391939) [SHA512-based CRYPT]... DONE
Raw:    49.7 c/s real, 49.7 c/s virtual

(we might be able to bump the CUDA figure but I don't have time for
that. Ideally CUDA formats should auto tune like the OpenCL formats do)

Using device 0: GeForce GTX 570
Benchmarking: crypt SHA-512 (rounds=391939) [OpenCL]... DONE
Raw:    4185 c/s real, 4185 c/s virtual

Using device 0: Tahiti
Benchmarking: crypt SHA-512 (rounds=391939) [OpenCL]... DONE
Raw:    5285 c/s real, 3276K c/s virtual

Using device 1: AMD FX(tm)-8120 Eight-Core Processor
Benchmarking: crypt SHA-512 (rounds=391939) [OpenCL]... DONE
Raw:    1873 c/s real, 234 c/s virtual

This last test should confirm that something's not right. I believe
OpenCL CPU can't be 70x OMP...

Whatever is wrong with Claudio's format, it's not caught by the
self-tests. We need to implement cryptsha in the test suite :(

> Also, I guess you were not running Claudio's latest OpenCL code, which
> should do 11.5k c/s on GTX 570 at rounds=5000.

It turns out I was. Not sure what is happening here.


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.