Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Sat, 25 Apr 2015 21:57:47 +0300
From: Solar Designer <solar@...nwall.com>
To: john-dev@...ts.openwall.com
Subject: Re: [GSoC] JtR SIMD support enhancements

On Sat, Apr 25, 2015 at 08:44:26PM +0200, magnum wrote:
> On 2015-04-25 20:21, Solar Designer wrote:
> >Speaking of OpenMP in particular, as you know it's inefficient for fast
> >hashes, so for those I'd tune the interleave factors with -fork instead.
> 
> Since we currently can't have different interleaving factors for 
> different formats using same primary hash function, I think it's most 
> sensible to simply tune for md5rypt, wpapsk and so on. I have seen a few 
> cases where raw-md5 and md5crypt had different best factor, but it was 
> insignificant - and md5crypt is the arbitrator IMO.
> 
> For MD4 we might actually want to tune for raw-md4 or NT2 though, using 
> fork like you say.

I agree.

> I guess Lei compiled with --disable-openmp-for-fast-formats, virtually 
> all formats can do OpenMP now even though it's not always sensible.

No, I think Lei didn't use that option.  I saw plenty of dynamic*
showing 240xOMP and poor speeds there (as expected).

> BTW 
> even for fast formats I have a feeling we might want to compare 59 forks 
> x 4 threads with simply 236 threads. Wouldn't that have better chance to 
> run optimally? Though maybe you've already said that at some point and 
> that's why I get this gut feeling :-)

I recall experimenting with this when I ran descrypt on our 5110P during
some contests.  I don't recall which split turned out to be optimal, and
it'd vary by hash type anyway.  Overall, yes, this is something to try
as well.

And it's 60 and 240, not 59 and 236.  One core is wasted only with
OpenMP offload and OpenCL, whereas we're primarily doing full native
OpenMP builds now.  Lei only managed to get bcrypt sort of work with
OpenMP offload so far, and I hope that's in a separate branch.

Alexander

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.