Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20120314155129.GA4166@openwall.com>
Date: Wed, 14 Mar 2012 19:51:29 +0400
From: Solar Designer <solar@...nwall.com>
To: john-users@...ts.openwall.com
Subject: Re: Best performance MPI vs OMP

On Thu, Mar 08, 2012 at 11:40:18AM +0100, Javier Gonz?lez del T?nago Liberal wrote:
> DES (many/one salt)
>     - MPI: 54410K/51601K
>     - OMP: 47087K/21626K
> 
> LM
>     - MPI: 786762K
>     - OMP: 50823K (1)
...
> (1) With OMP_NUM_THREADS=24

These look reasonable to me.  As discussed before, OpenMP
parallelization for LM hashes is currently very inefficient.

> NT
>     - MPI: 533520K
>     - OMP: 18498K (2)

With --format=nt, there's no OpenMP parallelization at all.  With
--format=nt2, there's an attempt at it, but it's even worse than LM's.

> NETLM (many/one salt)
>     - MPI: 53168K/17572K
>     - OMP: 27787K/865569 (1)
> 
> NETNTLM (many/one salt)
>     - MPI: 53268K/39016K
>     - OMP: 30736K/7077

OK.  (I think you lost a "K" on the last line, though.)

> I did real test and these are the results:
> 
> Crack LM Hash of "ex&31"
>     - MPI: 1m
>     - OMP: 29m
> 
> Crack LM Hash of "ex&314"
>     - MPI: 1m
>     - OMP: 55m
> 
> Crack LM Hash of "gol27$"
>     - MPI: 27sec
>     - OMP: 11m

OK.

> With NT Hash the results are quite similar, there is a big difference 
> between MPI and OMP.

You mean, between MPI and no parallelization.

> I will publish the results on the wiki.

I see that you did so now - thanks! - but I'd also like to see the more
relevant MD5-crypt and bcrypt results - can you please add those as well?
(I say that they are more relevant because they actually have reasonable
OpenMP parallelization efficiency, unlike LM with its known poor
efficiency and NT with its lack of OpenMP parallelization.  Another
reason is that it is more realistic to have strong passwords with them,
hence auditing them is more reasonable.)

Meanwhile, a friend sent me benchmark results for 4x X7560 (64 threads),
which I've just added to the wiki:

http://openwall.info/wiki/john/benchmarks

The best ones are (for clean 1.7.9, no jumbo):

GOMP_SPINCOUNT=1000000:

Benchmarking: Traditional DES [128/128 BS SSE2-16]... DONE
Many salts:     73098K c/s real, 1142K c/s virtual
Only one salt:  26699K c/s real, 416956 c/s virtual

Benchmarking: FreeBSD MD5 [32/64 X2]... DONE
Raw:    344603 c/s real, 5385 c/s virtual

Benchmarking: OpenBSD Blowfish (x32) [32/64 X2]... DONE
Raw:    23500 c/s real, 367 c/s virtual

OMP_NUM_THREADS=32

Benchmarking: Traditional DES [128/128 BS SSE2-16]... DONE
Many salts:     61675K c/s real, 1927K c/s virtual
Only one salt:  31241K c/s real, 976342 c/s virtual

I am also asking him for similar results for -jumbo-5 built as -x86-64i
(should have much better FreeBSD MD5 speed).

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.