|
Message-ID: <20161124183523.GA8292@openwall.com> Date: Thu, 24 Nov 2016 19:35:23 +0100 From: Solar Designer <solar@...nwall.com> To: john-users@...ts.openwall.com Subject: Re: Unexpected speeds with o5logon and o5logon-opencl Hi Kevin, On Thu, Nov 24, 2016 at 12:43:31AM +0000, Kevin Miller wrote: > I'm unsure if I'm the problem or have some broken installation but I would > expect 3 GPU's to surpass the speed of 6-core CPU although it's quite the > opposite in my case. > > When running the o5logon-opencl format the GPU's do not seem utilize and > appear idle while going through the process, the CPU is quicker! Unfortunately, yes. JtR's GPU support is still mostly limited to so-called "slow hashes". For some "fast hashes", we have on-GPU mask mode, which makes them in fact run fast. (And our mask mode can be combined with other modes, to produce hybrid modes.) Unfortunately, our o5login-opencl lacks on-GPU mask mode currently, yet this is a fast hash. So you can't even remedy the problem by using mask mode, as you would with some other fast hashes. (You're currently trying to run wordlist with rules, which would be inefficient with any fast hash.) You may improve things a little bit by using each GPU more than once, but it will still be very inefficient. Your original command: ./john --devices=0,1,2 --rules=best64 --format=o5logon-opencl --wordlist=passwords.txt o5logon_hash.txt (BTW, it used only 1 GPU since you didn't specify "--fork=3".) Revised for two processes per GPU: ./john --devices=0,1,2 --fork=6 --rules=best64 --format=o5logon-opencl --wordlist=passwords.txt o5logon_hash.txt or even for four processes per GPU: ./john --devices=0,1,2 --fork=12 --rules=best64 --format=o5logon-opencl --wordlist=passwords.txt o5logon_hash.txt On CPU, you can also speed things up a little bit by using "--fork" instead of OpenMP. Your original command: ./john --rules=best64 --format=o5logon --wordlist=passwords.txt o5logon_hash.txt Revised: ./john --fork=6 --rules=best64 --format=o5logon --wordlist=passwords.txt o5logon_hash.txt Also, you may choose to run wordlist+rules on CPU, and non-wordlist modes (incremental, mask, Markov) on GPU (actually, they run on CPU as well, but they compute the hashes on GPU). With wordlist+rules, the bottleneck of candidate password generation is particularly bad. > Below is some system info and examples to skim over. > > http://pastebin.com/raw/1f2PEXEh Thanks. On future occasions, please post such info directly to the list instead of via pastebin. 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.