Date: Thu, 8 Aug 2013 10:33:44 -0400 From: Milen Rangelov <gat3way@...il.com> To: john-users@...ts.openwall.com Subject: Re: Re: --fork using different OpenCL devices Yes, that is definitely a problem and the solution is not to distribute the work even (I break the keyspace into chunks and keep a big table that is updated by worker threads, this is ugly). Since the threads share the same address space, this is relatively easy, but in the forked processes case it may not be that easy. On Thu, Aug 8, 2013 at 9:06 AM, Claudio André <claudioandre.br@...il.com>wrote: > I think slow formats would benefit as well. [...] >> >> Regards, >> Milen >> > Hi. > Well, when I (somehow) tested it, I was using . > > The main idea is that each node will run in one OpenCL device. > Node 1 run using OpenCL device. > Node 2 run using OpenCL device. > And so on. > ---- > >  > $ ../run/john -form:sha512crypt-opencl -dev=1,2 ~/testhashes #It > translates to --fork=2; node=1->dev=1, node=2->dev=2 > ... > Local worksize (LWS) 64, global worksize (GWS) 32768 > Loaded 30 password hashes with no different salts (sha512crypt-opencl, > crypt(3) $6$ [SHA512 OpenCL]) > Node numbers 1-2 of 2 (fork) > ... > Local worksize (LWS) 1, global worksize (GWS) 1024 ... > 1 15g 0:00:02:48 3/3 0.08897g/s 1433p/s 1433c/s 31337C/s abdesy..sevoty > 2 15g 0:00:03:44 3/3 0.06668g/s 1808p/s 1808c/s 32620C/s ciaga1..223qra > ---- > > Magnum already pointed out some important details: > - Faster GPU(s) finish their work first. > - How should JtR handle it? > - Faster devices run faster, slower run slower. They do not impact each > other (different processes). Note: faster devices will have to wait slower > units finish their work. > - For fast hashes, maybe is better to focus on one device? All nodes > should send data to the same device? > > Claudio >
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.