Date: Mon, 7 Jan 2013 04:03:23 +0400 From: Solar Designer <solar@...nwall.com> To: john-dev@...ts.openwall.com Subject: des-opencl Sayantan - While posting comments here: http://www.reddit.com/r/crypto/comments/162ufx/research_project_opencl_bitslice_des_bruteforce/ I tested des-opencl currently in unstable-jumbo: OpenCL platform 1: AMD Accelerated Parallel Processing, 2 device(s). Using device 0: Tahiti Benchmarking: Traditional DES [OpenCL]... DONE Many salts: 43805K c/s real, 141669K c/s virtual Only one salt: 25022K c/s real, 41287K c/s virtual As we know, this is reasonable speed given no specialization for a salt value yet and given the formats interface bottleneck. However, when I went to test it on an actual password file, I found that: 1. Detection of matching salts does not work. This is easily fixed by changing BINARY_SIZE and SALT_SIZE in opencl_DES_fmt.c to be sizeof(WORD) (for consistency with your other changes) rather than ARCH_SIZE (a leftover from the CPU-based format). Please apply this change. I briefly tested it, and it works for me (fixes the problem, does not introduce side-effects). 2. The speed seen during actual cracking is a lot lower - about 1M c/s, not 44M c/s as --test would suggest (this is for many salts). Why is that? How to avoid this problem? A guess: you're using global memory somewhere, so caching helps when there are only a handful of different plaintext passwords/salts being hashed repeatedly. It could be something entirely different, though. Maybe you've recently made "optimizations" and did not test their effect other than on --test speeds, which do not always reflect real-world speeds. Maybe you need to revert some changes. When we were at around 20M c/s for --test, this was a speed we could actually achieve, right? I vaguely recall actually achieving it in a run on a password file. Thanks, 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.