Progress on the GPU accelerated RTs!

[chazchazam]

If you would like some more testing units I have quite a range.

Intel i7 - 2.66GHZ Speed step to 2.8
Intel i5 - 2.66GHZ Speed step to 2.8
AMD Phenom II - 2.3 GHZ

3x ATI 5870s
2x nvidia 295s
1x Nvidia 8800

Linux or Windows 7

Also should be acquiring a 5970 and at least one Nvidia 480 in the near future and plenty of storage.

(very generous University)

My friend and I are working on a research paper on GPGPU Computing and their effect on password cracking, This project could be a big help for our research.

[Bitweasil]

Excellent.

Sorry about the site outage. Should be a bit more stable now... unless I'm running table merges or something and blowing out the filesystem cache. I'll try not to do that *too* often.

Anyway, I've got the full process GPU accelerated and GUI friendly now, for single GPUs.

~40s to crack a 8 character loweralpha password in testing with chain lengths of 100k and a GTX260, and the table in memory (filesystem cache).

... right now, the code has some hardcoded MD5 len8 assumptions I'm fixing, and I've got a few small fixes to make (things like the table generate code being RAM limited and not having options to iterate - build me 10 5M element tables).

And a few other tweaks related to being fully usable. Being able to throw the cracker at a directory full of files and have it test against all the proper files, for instance.

I'm hoping to have the work done by the weekend. Depends on a few other things - chances of it being done are much higher if it's bad weather, otherwise I'll be out motorcycling/hiking.

[Bitweasil]

So, theoretically, by tomorrow, I'll have a full set of length 7 MD5 full-us-charset tables ready to go.

Chain length: 200k
Chains per table: 100M
Tables: 4

Total size: ~12.8GB

Theoretically, this will work...

I have the generator code totally working for MD5 len6-10. I want to extend the lengths a bit, add NTLM support, and I think it'll be ready to let some other people play with!

[Bitweasil]

... as was kindly pointed out on FRT, my table size math was more than a bit off.

400k elements to cover a table with high probability of success does NOT mean you can break that into 4x100M tables and keep the same success probability. That's what I get for starting a generate task at 1AM.

More math on the horizon, and probably a slightly larger tableset.

[sapling]

Still can't wait for your beta code release! I was getting worried about the site outage thinking you had run into issues like funding or time and just trashed the whole project. Glad to see that is not the case. Are you expecting to perhaps release your beta code after this weekend? I know its wishful thinking but I can dream! lol

[Bitweasil]

Yeah, still working on it.

The server was one being used for other stuff and I didn't have access to it, so if it went down, I had to send a nice email and wait.

I can't promise beta code this weekend as the gf is over, but if I get some time, I may see what I can work up - it shouldn't take *too* long to refactor everything and at least release MD5 code. NTLM will be coming soon, but I need to write more to support that.

[1epi]

if you want, i can give you some suggestions on the rainbow table characteristics. (i'm _haxxor_ )
just give me some details, what size, the speed of generation (if you can benchmark) and the success probability you want. (though 99.9% is more than enough, miss 1 in 1000 ), don't forget the charset.

[Bitweasil]

Ah, excellent. I will definitely take you up on that.

For now, let's assume Len 7 and Len 8, full US charset.

I can generate at around 400M MD5 links/sec and 600M NTLM links/sec, give or take. Per GT200 core.

Chain storage in the tables takes 32 bytes per chain + 8192 bytes of table header.

Precomputation for Len100k chains takes about 15s, Len200k takes about 60s (again, single GT200 core), then table search takes a while and chain regen takes a while (though it is GPU accelerated now).

My table merging code can either create perfect or non-perfect tables. Or both, should one care.

Anything else you'd find useful?

[Bitweasil]

Another bit of data:

95^7 password space, len200k chains, 400000000 total chains, 41.2% merge rate when perfecting the tables (41.2% of chains are duplicates).

[1epi]

Ah, excellent. I will definitely take you up on that.

For now, let's assume Len 7 and Len 8, full US charset.

I can generate at around 400M MD5 links/sec and 600M NTLM links/sec, give or take. Per GT200 core.

Chain storage in the tables takes 32 bytes per chain + 8192 bytes of table header.

Precomputation for Len100k chains takes about 15s, Len200k takes about 60s (again, single GT200 core), then table search takes a while and chain regen takes a while (though it is GPU accelerated now).

My table merging code can either create perfect or non-perfect tables. Or both, should one care.

Anything else you'd find useful?

full US charset = mixalpha-numeric-symbol32-space ?
charSetLen = 95 ?