[unpackme] Armadillo v6.0.4 Custom Build

[ghandi]

Hmmm, not allowed to edit my post, edit button has disappeared? Anyway, i've played with the concurrency by launching 2 kernels per host thread (4 host threads) that use separate streams, this way the first DeviceToHost memcpy is performed as the second calculation takes place. This has the result of pushing the total bruteforce time for older Armadillo checksums down to 4.85 seconds, which averaged 1,073,741,824 custom md5 operations in that time .

Strangely enough, now looking in the CUDA profiler, it says the kernels are leaving some multiprocessors idle... If i can figure out how to make them all work together with this new stream approach i wonder how low the total time can be pushed... This is important because the newer checksum method is 1000 times more work for the bruteforcer... I also wonder at the wisdom of using 4 threads to access the GPU as opposed to a main accessor thread which dispatches work to other threads to process after getting results from the GPU...

Again, i have not looked at the newer method i have outlined in above posts.

HR,

Ghandi

Edited April 28, 2012 by ghandi

[mrexodia]

@Ghandi: Wow, those are very good results... The bruter I got (not coded by me) only reaches 429,496,730 hashes per second, will try to implement your tips to make it faster..

Greetings,

Mr.eXoDia

[ghandi]

Newer bruteforcing is working properly but still projects 2h:16m for completion. I am wondering about creating threads to run indefinitely on the GPU, using atomic operations and checking for a result with async memcpy. There is a tradeoff, if i include code to validate the potential symmetric key(s) found, it will reduce the amount of memory available thus reducing the amount of threads i can launch on the device.

I do not know if this is worth it or even if i can make it work, i would hope that at least with older symmetrics i can make it so the bruteforcer spits out the one valid symmetric instead of potential multiples which need to be tested before the 'good' one is found.

HR,

Ghandi

[Apuromafo]

interesting noticia, maybe , in the new researching must work as say in old time togeder with mr.exodia and ghandi., both are good coders and crackers ^^ for solve the new armadillo

BR, APuromafo