Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20120807213153.GA21069@openwall.com>
Date: Wed, 8 Aug 2012 01:31:53 +0400
From: Solar Designer <solar@...nwall.com>
To: john-dev@...ts.openwall.com
Subject: Re: Result of hard core password generation on 7970

On Wed, Aug 08, 2012 at 02:56:54AM +0800, myrice wrote:
> On Wed, Aug 8, 2012 at 1:40 AM, Solar Designer <solar@...nwall.com> wrote:
> > On Tue, Aug 07, 2012 at 11:46:02PM +0800, myrice wrote:
> >> 1
> >> guesses: 0  time: 0:00:01:26 0.00%  c/s: 1956M
> >>
> >> 1000
> >> guesses: 0  time: 0:00:01:27 0.00%  c/s: 1807G = 1807M
> >>
> >> 1M
> >> guesses: 0  time: 0:00:01:25 0.00%  c/s: 922618G = 922.6M
> >
> > This certainly looks better, but where is this code?  I am trying out
> > PG-test now, and it is not giving these speeds.
> 
> Sorry, I remember I have updated to my github, but it's not. Now it is here

Thanks.  Attached is a patch with some optimizations and comments about
further optimizations to make (perhaps after you're done with bitmaps).

Your direct comparisons, as implemented, turned out to be slower than
local memory bitmap lookups, so I #if 0'ed them out here.  If we choose
to implement them, then maybe try it for the case of exactly 1 loaded
hash only, and avoid the array and loop.

1
guesses: 0  time: 0:00:01:34 0.00%  c/s: 2083M

1000
guesses: 0  time: 0:00:01:28 0.00%  c/s: 1941G

These are a bit faster now.

1M
guesses: 0  time: 0:00:01:26 0.00%  c/s: 851953G

No speedup for 1M - in fact, a slowdown on this specific test as
compared to your run, but that's kind of random.  Indeed, this test is
limited by global memory latency.

What has happened to the idea of using 4 x 128 MiB bitmaps (like
Multiforcer does) before the hash table lookup (no need to implement
binary search to try this out when we already have the hash table)?
I thought this was on your priorities for this week.

Thanks,

Alexander

View attachment "john-PG-test-opencl-rawmd5-opt1.diff" of type "text/plain" (11484 bytes)

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.