Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Wed, 8 Jun 2011 00:10:16 +0400
From: Solar Designer <>
Subject: Re: Lukas's Status Report - #4 of 15


On Tue, Jun 07, 2011 at 09:46:18PM +0200, ?ukasz Odzioba wrote:
> I've implemented md5-based crypt but still having trouble with forcing
> it to work properly.

Maybe you can show your code, even if non-working?  There might be
design changes that I'd propose (regardless of implementation bugs).

> In the meantime I found cpu implementation called "fast md5-crypt" and

Can you please post an URL for it (or otherwise share it with us)?
A Google web search for "fast md5-crypt" doesn't appear to find it.

> want to analyze it and compare to the current one.

You mean to the one in JtR?  BTW, bartavelle's code in 1.7.7-jumbo-5
introduces a SIMD implementation.  Perhaps you want to look at non-SIMD
code in official 1.7.7.

Also, since the overall structure of SHA-crypt was based on MD5-crypt's,
you may be able to reuse some approaches for SHA-crypt.

> I tried to check coalesced read/write in sha256 crypt (because it was
> easiest to do) and got following results:
> Uncoalesced
> Raw:    10569K c/s real, 10569K c/s virtual
> Coalesced read
> Raw:    12115K c/s real, 12115K c/s virtual
> +4.6%
> Coalesced read and write
> Raw:    12376K c/s real, 12376K c/s virtual
> +2.2%
> Overall: +7%

OK.  Obviously, the bottleneck for fast hashes like this is different.

> It is not a big deal, but it's worth to do in other patches.

Sounds right.

> Surprisingly I didn't stated any difference in coalescing operations
> inside kernel.

What do you mean?  Perhaps you wanted to say "notice", not "stated", but
even if so I find this unclear.

> I have also added gtx460 benchmarks results to the wiki.  Presenting
> them as a list its not user friendly, so meaby we could turn on html
> or php support on wiki to add charts as a nice google chart script,
> example here:

The chart looks pretty, but I have no problem with textual data on the
wiki so far.  I think we should focus on actually improving the code.
So far, for everything but phpass (and the artificial looped SHA-256
that I suggested as an experiment) the performance is very poor.  But I
trust you'll be able to improve it.



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.