Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date: Wed, 17 Jul 2013 03:20:08 +0530
From: Sayantan Datta <>
Subject: Re: Sayantan's weekly report #5

On Wednesday 17 July 2013 12:08 AM, Solar Designer wrote:
> Hi Sayantan,
> On Mon, Jul 15, 2013 at 08:16:54PM +0530, Sayantan Datta wrote:
>> Accomplishments:
>> - mask mode for Raw-MD5-opencl.
> Great.  Is it normal (for your current tree) that descrypt-opencl and
> raw-md5-opencl fail --test (in fact, the former segfaults) and that only
> --mask mode works (no other modes)?

Yes the --test options are failing. However both the formats performs 
self test before cracking. Even if I can make them pass the --test 
options it would not give correct benchmark numbers.
Currently no other modes are supported. Let me first iron out all the 
bugs with mask mode, then I will integrate the other modes.

>> Priorities:
>> -  debug the Raw-MD5-opencl password generator, there are some deeper
>> bugs(could be compiler) that is causing wrong password to be generated in
>> some cases.
>> -  Optimize Raw-MD5-opencl. Current max speed is nearly 3100 M c/s
> I've experimented with it a little.  It successfully cracked my test
> password (given a proper mask).  The reported speeds appear to be wrong.
> I am getting:

I am not updating the *pcount number because when *pcount exceeds 
max_keys_per_crypt ,it causes segfault and I'm yet to find out where.

> solar@...l:~/j/s/JohnTheRipper/run$ time ./john -form=raw-md5-opencl pw -mask=?l?l?l?l?l?l?l?l
> Device 0: Tahiti (AMD Radeon HD 7900 Series )
> Local worksize (LWS) 64, global worksize (GWS) 4194304
> Loaded 1 password hash (Raw-MD5-opencl [MD5 OpenCL (inefficient, development use only)])
> [...]
> 0g 0:00:01:02 0g/s 4944Kp/s 4944Kc/s 4944KC/s aoyuhtur..CHECK
> Session completed
> real    1m19.917s
> user    1m14.749s
> sys     0m59.696s
> However, 26^8/62 = 3368178460, which is 3368M, not 4944K.
> Also, why does it spend 17 seconds on initialization - what is it doing
> during that time?  Note: most of these 17 seconds are after the Loaded
> line appears, so it's not LWS/GWS tuning.

  The initialization part for raw-md5 opencl was done by claudio and 
magnum. I think it is performing self test in those 17 secs. Also the 
auto tuning isn't accurate for the cracking kernel, producing wrong LWS 
and GWS size. Probably the auto tuning needs to be done in reset for 
mask mode.

>> -  Start a new format. Please discuss!!
> I think NTLM, MSCash (DCC), LM may be next.  (You don't have LM on GPU
> yet, right?  So you'd need to implement it along with mask right away.)
> Also, how about making your code closer to becoming suitable for merging
> into the main bleeding-jumbo tree?  What would you need to do to make a
> merge possible?

I would start with integrating the mask-mode interface with bleeding. 
Currently the formats lack auto tuning and I need to figure out where 
and how to do that. Also some of the segfault issues needs to be 
addressed before we can merge into main bleeding.

>> -  write experimental isa code using Daniel's gcnasm.
> This could help Daniel's project, so I would in fact appreciate it if
> you give this a try - just don't let it result in your own project
> getting stuck/abandoned.  Thanks!
>> git clone -b
>> bleeding-jumbo
> Alexander

Yes, of course.


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.