Date: Tue, 1 May 2012 08:37:02 +0400 From: Solar Designer <solar@...nwall.com> To: john-dev@...ts.openwall.com Subject: Re: Lukas - status report #2 Lukas - On Tue, May 01, 2012 at 06:13:01AM +0200, Lukas Odzioba wrote: > 2012/5/1 Solar Designer <solar@...nwall.com>: > > http://openwall.info/wiki/john/WPA-PSK > > > > Maybe you can add a specific example to it (a sample input file, > > commands to run on it, their output) and link to it more prominently? > > Of course I'll add example. > Have you got any suggestions how to make it more prominently? Move it > on top of page, or link in gpu formats table? Link from the GPU page, link from the tutorials page. Maybe even have a separate line for links to non-hashes pages on the main /wiki/john page. > > Here's what I am getting with the code currently in magnum-jumbo: > > > > user@...l:~/john/magnum-jumbo/run$ ./john -te -fo=wpapsk-cuda > > Benchmarking: wpapsk-cuda [GPU]... DONE > > Raw: 17341 c/s real, 17341 c/s virtual > > > > This is 43% of hashcat's reported speed for this card. > > GTX460 with sm_20 and threads=256 does ~15k, 10k by default. Oh, I haven't tried any tuning yet. Please do (on bull) and add to doc/README-CUDA. And post about this to john-dev, indeed. > > user@...l:~/john/magnum-jumbo/run$ ./john -te -fo=wpapsk-opencl -pla=1 > > OpenCL platform 1: AMD Accelerated Parallel Processing, 2 device(s). > > Using device 0: Tahiti > > Optimal Group work Size = 128 > > Benchmarking: wpapsk-opencl [pbkdf2-sha1]... DONE > > Raw: 64000 c/s real, 531692 c/s virtual > > > > This is quite nice. hashcat is reported to do 158.1k c/s on 5970, so > > our target speed for 7970 may be about 130k c/s. > > I would be more happy to see 80-90k, previously (just pmk calculation > - most time consuming) we had 90% of hashcat's speed. For now > difference will be ever worst for super fast gpus and slow cpu. > Besides cpu side code utilizes only 1 core. Do you have any ideas to > get around it other than MPI? On the other side we could move all code > to second kernel gpu. I don't understand what you're referring to. I just took a look at opencl_wpapsk_fmt.c and I don't see it doing much on the CPU. Can you point me at specific places in the code? Do you think my run on the 7970 was somehow CPU-bound? I doubt it. > > user@...l:~/john/magnum-jumbo/run$ ./john -te -fo=wpapsk-opencl -pla=1 -dev=1 > > OpenCL platform 1: AMD Accelerated Parallel Processing, 2 device(s). > > Using device 1: AMD FX(tm)-8120 Eight-Core Processor > > Optimal Group work Size = 16 > > Benchmarking: wpapsk-opencl [pbkdf2-sha1]... DONE > > Raw: 2133 c/s real, 267 c/s virtual > > > > This is also reasonable, although a CPU-specific implementation using > > the intrinsics should be much faster. > > My current code is based on openssl (it's not yet in jumbo), and It > gives ~230 c/s on i3 2100, in Aircrack-ng i have 510 c/s (uses 4 > threads). > As far as I know openssl is not for super optimized, and with > intrisics we should get much better results, am I right? OpenSSL is optimized reasonably well for its interface, but the interface is not well-suited for password cracking. We have to make three calls to compute one SHA-1 hash, and the SHA1_Final() call probably wastes some time cleaning the "sensitive data" out from memory. What's more important, it only computes one hash at a time, whereas with the intrinsics we can do 4 at a time (on SSE2 or better), or even more with mixed instructions for greater instruction issue rates. So, yes, we should get much better results with the intrinsics. Am I correct that we compute SHA-1 8192 times per WPA password checked (4096 iterations, and two SHA-1's per HMAC)? If so, the 23M+ c/s at raw SHA-1 that I am getting with the XOP code on one CPU core would translate to over 2800 c/s at WPA. That's on one core. This does not include overhead to implement HMAC and PBKDF2, though. So maybe something between 5000 and 10000 c/s on this FX-8120 CPU is realistic. 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.