Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date: Fri, 22 Jul 2011 02:52:38 +0400
From: Solar Designer <>
Subject: testing of cryptmd5cuda (was: Lukas's Status Report - #10 of 15)

On Fri, Jul 22, 2011 at 12:24:23AM +0200, ?ukasz Odzioba wrote:
> I just did a simple test. But I'dont know is the methodology right. I
> just run ./john hashes.txt and watched c/s for 10 minutes.
> I'v got 9000 c/s on my i3-2100 (3.1Ghz 1thread) so, on QUad-core
> should be around 50k, which is only 5times slower than my gpu patch.
> Have you got any sugestions about methodology, which is probably not good.

This testing is primarily to make sure the patch works correctly.  The
c/s rate is also very important, but it is secondary (who cares how fast
the program runs if it doesn't work).

On Fri, Jul 22, 2011 at 12:25:33AM +0200, ?ukasz Odzioba wrote:
> I forgot to add "logs"
> cpu:
> gpu:

I took a look at these.  The CPU one (somehow a 32-bit build?!) loads
4716 hashes, cracks 211 passwords in a little over 10 minutes.  The GPU
one also loads 4716 hashes, but fails to crack any passwords in almost
12 minutes.  Yet when you run "--show" right after that GPU run, it
shows the 211 passwords previously cracked on CPU.

So the cryptmd5cuda patch has at least two problems:

1. It does not recognize passwords previously cracked by the CPU run
(they should have been excluded from loading, but they were not).

2. It fails to crack any passwords.

This makes me wonder how you have been testing it so far.  Just "--test"?

Please re-focus on making your patches actually work before you proceed
with OpenCL.



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.