john-users - Re: New NT patch

Openwall GNU/*/Linux 3.0 - a small security-enhanced Linux distro for servers

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date: Mon, 12 Mar 2007 23:11:39 +0100
From: "Alain Espinosa" <alainesp@...il.com>
To: john-users@...ts.openwall.com
Subject: Re: New NT patch

In the revision 5 are two different algoritms: generic and sse2

> 1- Now passwords are trying in blocks of lenght 40-64. This speedup the
> > performance (not in the benchmark, in the "real" c/s),particularly  in
> > the incremental mode with near a 30%.

This apply  to the 2 algorithms.

> 2- Add SSE2 code with a speedup of 20% over C code. Note that "plain"
> > SSE2 code are a little bit bad that C code because the emulation of
> > the rotate instruction that dont exist in SSE2 (gcc are smart enough
> > to produce rotate instruction in the C code). The speedup is because i
> > intermix SSE2 code with x86 code trying 5 passwords in parallel.
>
> This apply to the SSE2 code

> Is the large benchmark speed increase real?

Yes, its real. I put a 20% because i test in a Celeron D computer. If the
SSE2 implementation are better (and in the MacOSX with 2.33 Ghz
Core2 cpu. seen to be better) the speedup is more great.  And this was
compare with a C compiler that optimaly generate the code (puting rot
instructions).  1- improve the benchmark  too.

Try with incremental mode a bit to see what are the real speed. In my
computer with one hash and SSE2 algorithm:

benhmark:                    10400K
incremental:                   9000K

Generic algorithm:

benchmark:                  8600K
incremental:                 7400K
-- 

alain