[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <4E4E9083.3030507@bredband.net>
Date: Fri, 19 Aug 2011 18:34:11 +0200
From: magnum <rawsmooth@...dband.net>
To: john-dev@...ts.openwall.com
Subject: Re: 45% wordlist boost just waiting to happen
On 2011-08-19 14:42, JFoug wrote:
> From: "magnum" <rawsmooth@...dband.net>
>> I still see >40% boost for LM without nsk3 and >60% with nsk3. And now
>> passing test suite B-)
>
> Ok, if you are still getting 40% boost, with the code not chopping up
> the soon to be used words, then it is certainly worth getting it right.
>
> I will look at the code you sent me, along with the code I did to get
> full 'const ness' of the line, including making the call into
> rules.apply a const char *, and the misc.strcpy() stuff also using const
> char * where it should, and get a patch out.
I made a couple of changes. I just uploaded it as 0025, please rebase
onto that. I think it's best having these separate anyway, but feel free
to modify my version.
This patch is *thoroughly* tested, using -wordlist, -pipe, -stdin and
with/without memory buffer. Truncation and duplication suppression is
verified. Test Suite and a lot of manual tests passed. All tests on
linux-x86-64 only though (but I really don't expect problems on other
builds)
Performance boost remains (LM/nsk3 >60% on my Intel core duo) but varies
between machines. I saw only 27% boost for same test on an AMD machine I
tested on.
There is one caveat though: We are now truncating words as we load them
to memory. I'm not sure how things are supposed to work, maybe for a
length 27 format we should be able to load eg. a length 32 word, delete
5 characters from the *beginning* of it with a rule, and end up with the
remaining 27? If this is supposed to be supported we'll need to tweak it
a little.
magnum
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.
